diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_0/index_tensor([8076, 8076])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_0/index_tensor([8076, 8076])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..38dacf2ad9593d2d65a1ba87dbd9a119fb0f7208 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_0/index_tensor([8076, 8076])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up captures a perfect drop of shimmering golden maple syrup landing on a dark, matte, non-reflective surface. The droplet becomes a viscous, shimmering dome, its high-gloss surface reflecting a soft light, imbuing it with a warm, honey-hued glow. As the droplet slowly solidifies, a transformation begins almost imperceptibly; deep within its core, a tiny, star-shaped point of crystallization emerges. From this epicenter, a complex crystalline structure steadily and deliberately blooms outward, spreading like delicate frost flowers through the liquid gold. The once-smooth surface gives way to a complex network of geometric facets, light catching each new edge and causing it to shimmer. The process is a silent, mesmerizing ballet of molecular transformations, the viscous liquid meticulously reshaped, ultimately solidifying into a hard, translucent, multi-faceted amber gemstone that captures and refracts light in a dazzling prismatic display, creating an atmosphere of quiet, natural alchemy and focused wonder.", + "seed": 31063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_0/index_8076_seed_31063_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04788877507671714, + "vlm_logprob": -0.04788877507671714, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.04788877507671714, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012910117395222187 + ], + "opposite_logprob_tokens": [ + -4.887909889221191 + ], + "logprob_reward": -0.012910117395222187, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012910117395222187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09695641696453094, + 0.19614726305007935, + 0.0625680536031723, + 0.16273793578147888, + 0.07403910905122757, + 0.06940403580665588, + 0.09806867688894272, + 0.03957805782556534, + 0.15491826832294464, + 0.04558218643069267 + ], + "credit_map": [ + 0.009159416891634464, + 0.016583595424890518, + 0.007909075357019901, + 0.012527280487120152, + 0.011975127272307873, + 0.017433149740099907, + 0.010197406634688377, + 0.011171362362802029, + 0.024094412103295326, + 0.020148739218711853, + 0.02004462666809559, + 0.03539948910474777, + 0.017868679016828537, + 0.029102519154548645, + 0.026052629575133324, + 0.023436164483428, + 0.00794385839253664, + 0.0062619773671031, + 0.004928193055093288, + 0.008405151776969433, + 0.012420045211911201, + 0.00792528223246336, + 0.005123864859342575, + 0.00955967977643013, + 0.01503986120223999, + 0.02005857788026333, + 0.025864729657769203, + 0.018840644508600235, + 0.024061914533376694, + 0.020644966512918472, + 0.017617041245102882, + 0.020610202103853226, + 0.008425458334386349, + 0.0070007867179811, + 0.007786782458424568, + 0.010488144122064114, + 0.01228760089725256, + 0.00991042610257864, + 0.008892130106687546, + 0.009247777983546257, + 0.009501143358647823, + 0.006319769192487001, + 0.005233993753790855, + 0.006651916541159153, + 0.00771360844373703, + 0.019903650507330894, + 0.00680141756311059, + 0.007278540637344122, + 0.006957093719393015, + 0.01347796618938446, + 0.01010058168321848, + 0.0135439308360219, + 0.008408033289015293, + 0.02249077335000038, + 0.011070968583226204, + 0.012019341811537743, + 0.005449030548334122, + 0.006081078201532364, + 0.0041405982337892056, + 0.005468531046062708, + 0.004128227010369301, + 0.003496251767501235, + 0.005916396621614695, + 0.004897943697869778, + 0.013196993619203568, + 0.015713771805167198, + 0.021215448155999184, + 0.02199278585612774, + 0.024085702374577522, + 0.014823808334767818, + 0.016163107007741928, + 0.02772665023803711, + 0.004360448103398085, + 0.006654444150626659, + 0.006790987215936184, + 0.007026451639831066, + 0.006338678766041994, + 0.0037802595179528, + 0.005555063486099243, + 0.005075853783637285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.012910117395222187 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159615576267242 + ], + "opposite_logprob_tokens": [ + -1.7015961408615112 + ], + "logprob_reward": -0.20159615576267242, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159615576267242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16349875926971436, + 0.26812684535980225, + 0.05606038123369217, + 0.1411610245704651, + 0.04915577173233032, + 0.053928207606077194, + 0.08466575294733047, + 0.09019351750612259, + 0.04663487523794174, + 0.04657483845949173 + ], + "credit_map": [ + 0.02218271791934967, + 0.021208597347140312, + 0.02169143036007881, + 0.016391849145293236, + 0.01620616763830185, + 0.017135366797447205, + 0.021614495664834976, + 0.02706812135875225, + 0.035172201693058014, + 0.029255416244268417, + 0.022806093096733093, + 0.0437239371240139, + 0.02200924977660179, + 0.03507501259446144, + 0.029728688299655914, + 0.05035623908042908, + 0.009535055607557297, + 0.005807270761579275, + 0.005251375027000904, + 0.004571094643324614, + 0.01136570144444704, + 0.006820998154580593, + 0.004394405521452427, + 0.008314481005072594, + 0.017870109528303146, + 0.016968408599495888, + 0.019309936091303825, + 0.014695906080305576, + 0.025251595303416252, + 0.018446916714310646, + 0.014661752618849277, + 0.013956401497125626, + 0.004985923413187265, + 0.003674047999083996, + 0.005009693559259176, + 0.006057336460798979, + 0.0073790643364191055, + 0.006830192636698484, + 0.005988794378936291, + 0.009230717085301876, + 0.005193409975618124, + 0.0057529546320438385, + 0.006011038552969694, + 0.007323897909373045, + 0.007126626558601856, + 0.009357455186545849, + 0.007007503416389227, + 0.006155322305858135, + 0.008733698166906834, + 0.01239401288330555, + 0.00917806476354599, + 0.008315312676131725, + 0.008647022768855095, + 0.01230758335441351, + 0.011730148456990719, + 0.01335990335792303, + 0.015343702398240566, + 0.029727570712566376, + 0.005396542139351368, + 0.007260813843458891, + 0.005489632021635771, + 0.017757276073098183, + 0.004091556183993816, + 0.005126419942826033, + 0.003917275462299585, + 0.005009595770388842, + 0.006295308470726013, + 0.006916534621268511, + 0.0045616282150149345, + 0.0038681577425450087, + 0.005427286494523287, + 0.010639090090990067, + 0.004591909237205982, + 0.00609313091263175, + 0.004855629988014698, + 0.007059314288198948, + 0.007649847771972418, + 0.005259414669126272, + 0.006143609527498484, + 0.004921985324472189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.20159615576267242 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064496710896492 + ], + "opposite_logprob_tokens": [ + -4.141064643859863 + ], + "logprob_reward": -0.016064496710896492, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064496710896492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09651261568069458, + 0.10589320212602615, + 0.11628345400094986, + 0.08473672717809677, + 0.18432186543941498, + 0.03739525377750397, + 0.12657569348812103, + 0.11857575923204422, + 0.08528796583414078, + 0.04441743344068527 + ], + "credit_map": [ + 0.010168187320232391, + 0.012186095118522644, + 0.013825262896716595, + 0.014598547481000423, + 0.009501510299742222, + 0.01611337624490261, + 0.010792111046612263, + 0.009327522478997707, + 0.01752397045493126, + 0.009416581131517887, + 0.0097111277282238, + 0.00824787188321352, + 0.007300893310457468, + 0.014966757036745548, + 0.008807586506009102, + 0.029918408021330833, + 0.014588359743356705, + 0.015740282833576202, + 0.013019800186157227, + 0.022359265014529228, + 0.015156307257711887, + 0.013384716585278511, + 0.008730724453926086, + 0.013304009102284908, + 0.012867745943367481, + 0.008637983351945877, + 0.008073480799794197, + 0.010149888694286346, + 0.011353296227753162, + 0.009976886212825775, + 0.009853671304881573, + 0.013823776505887508, + 0.01621405966579914, + 0.020695334300398827, + 0.018172120675444603, + 0.02495337836444378, + 0.03338102996349335, + 0.02611243911087513, + 0.02499694563448429, + 0.019796548411250114, + 0.004191006533801556, + 0.006008695811033249, + 0.003729506628587842, + 0.005168904550373554, + 0.005243684630841017, + 0.004143749829381704, + 0.005174195859581232, + 0.0037355099339038134, + 0.015438838861882687, + 0.023652920499444008, + 0.011198119260370731, + 0.017269035801291466, + 0.015452158637344837, + 0.014838054776191711, + 0.012293069623410702, + 0.01643349416553974, + 0.0215014535933733, + 0.03956414386630058, + 0.006667938083410263, + 0.010580187663435936, + 0.00710910651832819, + 0.01907953992486, + 0.008668449707329273, + 0.005404935218393803, + 0.00947276409715414, + 0.007298989687114954, + 0.010011225938796997, + 0.011404572986066341, + 0.010050054639577866, + 0.007995699532330036, + 0.010434446856379509, + 0.018620217218995094, + 0.004773262422531843, + 0.011164573021233082, + 0.0032707699574530125, + 0.006797338370233774, + 0.005831446964293718, + 0.0032787153031677008, + 0.005464229267090559, + 0.003837098367512226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.016064496710896492 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254624877125025 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.005254624877125025, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254624877125025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07136569917201996, + 0.11714909970760345, + 0.05282997712492943, + 0.1670033037662506, + 0.09999771416187286, + 0.032981351017951965, + 0.10618192702531815, + 0.1775352507829666, + 0.11635689437389374, + 0.058598801493644714 + ], + "credit_map": [ + 0.006137314718216658, + 0.011944353580474854, + 0.009521383792161942, + 0.009630625136196613, + 0.0061653549782931805, + 0.011824400164186954, + 0.007991822436451912, + 0.008150439709424973, + 0.018909446895122528, + 0.011569608002901077, + 0.011160156689584255, + 0.015841586515307426, + 0.008294586092233658, + 0.015600152313709259, + 0.00845054816454649, + 0.027323009446263313, + 0.007772640325129032, + 0.0067495242692530155, + 0.005839552264660597, + 0.008658364415168762, + 0.006530236918479204, + 0.006504300981760025, + 0.00384356826543808, + 0.006931792479008436, + 0.024589141830801964, + 0.016350675374269485, + 0.017889544367790222, + 0.02074016071856022, + 0.02431856468319893, + 0.01967741549015045, + 0.019526800140738487, + 0.023910999298095703, + 0.014138249680399895, + 0.008035986684262753, + 0.01187997404485941, + 0.010928076691925526, + 0.014317833818495274, + 0.01144923735409975, + 0.009047883562743664, + 0.020200472325086594, + 0.003355463268235326, + 0.0048415083438158035, + 0.004227325785905123, + 0.00417354516685009, + 0.004658651538193226, + 0.004640510771423578, + 0.003869858104735613, + 0.003214488038793206, + 0.013940633274614811, + 0.021338671445846558, + 0.00968919973820448, + 0.012059000320732594, + 0.012514817528426647, + 0.012052252888679504, + 0.010425192303955555, + 0.014162158593535423, + 0.031682342290878296, + 0.06117284670472145, + 0.010129272937774658, + 0.01311939861625433, + 0.010397948324680328, + 0.032512057572603226, + 0.009552949108183384, + 0.008968436159193516, + 0.012435337528586388, + 0.010786379687488079, + 0.0144357830286026, + 0.017493510618805885, + 0.013761199079453945, + 0.010591371916234493, + 0.015128541737794876, + 0.02172476053237915, + 0.005109732039272785, + 0.009252306073904037, + 0.004694996401667595, + 0.009666617959737778, + 0.010659047402441502, + 0.004720163065940142, + 0.008981546387076378, + 0.005514387972652912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.005254624877125025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036184806376695633 + ], + "opposite_logprob_tokens": [ + -5.6286187171936035 + ], + "logprob_reward": -0.0036184806376695633, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036184806376695633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05701413005590439, + 0.12630443274974823, + 0.036897990852594376, + 0.17185372114181519, + 0.11207789182662964, + 0.15377745032310486, + 0.10930559039115906, + 0.06067769601941109, + 0.08113071322441101, + 0.09096041321754456 + ], + "credit_map": [ + 0.007269682362675667, + 0.007135422900319099, + 0.006112990900874138, + 0.005758530925959349, + 0.006901861168444157, + 0.007935955189168453, + 0.006529004778712988, + 0.009370682761073112, + 0.02007044292986393, + 0.012317925691604614, + 0.010703478008508682, + 0.014436902478337288, + 0.010252674110233784, + 0.018521416932344437, + 0.011759151704609394, + 0.028242429718375206, + 0.004699842073023319, + 0.004193685948848724, + 0.004185827448964119, + 0.007369781844317913, + 0.004615504294633865, + 0.004490159917622805, + 0.0029101634863764048, + 0.004433026071637869, + 0.024873608723282814, + 0.017952635884284973, + 0.02013472653925419, + 0.02135320194065571, + 0.025681719183921814, + 0.01915706694126129, + 0.01968712918460369, + 0.02301362343132496, + 0.015194426290690899, + 0.008590197190642357, + 0.013711787760257721, + 0.01363898254930973, + 0.014370422810316086, + 0.013780047185719013, + 0.010784714482724667, + 0.02200731448829174, + 0.019755836576223373, + 0.021413668990135193, + 0.009998883120715618, + 0.018221130594611168, + 0.018654150888323784, + 0.03315186873078346, + 0.018322881311178207, + 0.014259041287004948, + 0.012460445053875446, + 0.01324863824993372, + 0.011437149718403816, + 0.013469316996634007, + 0.00951917003840208, + 0.01616698130965233, + 0.01422007568180561, + 0.01878381334245205, + 0.010260826908051968, + 0.017795773223042488, + 0.0038739792071282864, + 0.007134294603019953, + 0.0040296888910233974, + 0.007562173064798117, + 0.006344859022647142, + 0.0036760957445949316, + 0.006810765713453293, + 0.007606436964124441, + 0.01168770156800747, + 0.012574278749525547, + 0.009664531797170639, + 0.007169763557612896, + 0.01010589487850666, + 0.015511340461671352, + 0.007055347319692373, + 0.017367277294397354, + 0.010964528657495975, + 0.012567452155053616, + 0.015232020057737827, + 0.007654431276023388, + 0.012806909158825874, + 0.007312439847737551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.0036184806376695633 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures a perfect drop of shimmering golden maple syrup landing on a dark, matte, non-reflective surface. The droplet becomes a viscous, shimmering dome, its high-gloss surface reflecting a soft light, imbuing it with a warm, honey-hued glow. As the droplet slowly solidifies, a transformation begins almost imperceptibly; deep within its core, a tiny, star-shaped point of crystallization emerges. From this epicenter, a complex crystalline structure steadily and deliberately blooms outward, spreading like delicate frost flowers through the liquid gold. The once-smooth surface gives way to a complex network of geometric facets, light catching each new edge and causing it to shimmer. The process is a silent, mesmerizing ballet of molecular transformations, the viscous liquid meticulously reshaped, ultimately solidifying into a hard, translucent, multi-faceted amber gemstone that captures and refracts light in a dazzling prismatic display, creating an atmosphere of quiet, natural alchemy and focused wonder.", + "seed": 31064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_0/index_8076_seed_31064_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02560517154633999, + "vlm_logprob": -0.02560517154633999, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02560517154633999, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0435522198677063 + ], + "opposite_logprob_tokens": [ + -3.4185521602630615 + ], + "logprob_reward": -0.0435522198677063, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0435522198677063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2846863567829132, + 0.2046651989221573, + 0.07015460729598999, + 0.03999338671565056, + 0.05446768179535866, + 0.11820723861455917, + 0.12726180255413055, + 0.02813696675002575, + 0.04795386642217636, + 0.024472912773489952 + ], + "credit_map": [ + 0.03263925760984421, + 0.03132990747690201, + 0.03748142346739769, + 0.03814610466361046, + 0.029397491365671158, + 0.06053617224097252, + 0.022759508341550827, + 0.03239646926522255, + 0.01771962270140648, + 0.005242012441158295, + 0.010250459425151348, + 0.007271755952388048, + 0.005345168989151716, + 0.07546011358499527, + 0.006257528439164162, + 0.07711853086948395, + 0.0073276241309940815, + 0.0067451284267008305, + 0.009808352217078209, + 0.00817874539643526, + 0.009222578257322311, + 0.00886286236345768, + 0.010979976505041122, + 0.009029339998960495, + 0.004448682069778442, + 0.003190661547705531, + 0.006632387638092041, + 0.00377997406758368, + 0.0031131096184253693, + 0.0063841803930699825, + 0.005002386402338743, + 0.007442005909979343, + 0.006958243437111378, + 0.00539055559784174, + 0.007892479188740253, + 0.005793450865894556, + 0.004736622329801321, + 0.006234936416149139, + 0.009506830014288425, + 0.007954562082886696, + 0.012144690379500389, + 0.012601878494024277, + 0.02512121945619583, + 0.012194504961371422, + 0.007637279573827982, + 0.0153872761875391, + 0.017884189262986183, + 0.015236200764775276, + 0.010193264111876488, + 0.012833678163588047, + 0.02437123842537403, + 0.024401990696787834, + 0.009043228812515736, + 0.013769331388175488, + 0.01824893243610859, + 0.014400145970284939, + 0.003822346217930317, + 0.0023325884249061346, + 0.005305672530084848, + 0.002457256428897381, + 0.0031919258181005716, + 0.005158668849617243, + 0.0023995451629161835, + 0.003468962386250496, + 0.004512517247349024, + 0.009699502028524876, + 0.005472242832183838, + 0.0053147682920098305, + 0.006445039063692093, + 0.007127059623599052, + 0.004934885073453188, + 0.004447850864380598, + 0.003077463945373893, + 0.0023129780311137438, + 0.005091689061373472, + 0.0033461886923760176, + 0.0017397744813933969, + 0.003386308439075947, + 0.0027888766489923, + 0.002729634987190366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0435522198677063 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207595765590668 + ], + "opposite_logprob_tokens": [ + -2.8120758533477783 + ], + "logprob_reward": -0.06207595765590668, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207595765590668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5815989971160889, + 0.08126702904701233, + 0.09922115504741669, + 0.02038637362420559, + 0.03402863070368767, + 0.07735858112573624, + 0.018635688349604607, + 0.036703843623399734, + 0.023997806012630463, + 0.02680191956460476 + ], + "credit_map": [ + 0.06898608058691025, + 0.03621871769428253, + 0.1299980729818344, + 0.06524934619665146, + 0.033118076622486115, + 0.10102451592683792, + 0.047551143914461136, + 0.09945302456617355, + 0.011449549347162247, + 0.004068281501531601, + 0.022314175963401794, + 0.007428702898323536, + 0.005965434014797211, + 0.014370426535606384, + 0.007678607478737831, + 0.007991843856871128, + 0.010546287521719933, + 0.00909304991364479, + 0.01439721416682005, + 0.011708087287843227, + 0.014046239666640759, + 0.01292586699128151, + 0.016983643174171448, + 0.009520755149424076, + 0.0031318352557718754, + 0.0018349757883697748, + 0.003669033758342266, + 0.002142092678695917, + 0.002107305685058236, + 0.0028568897396326065, + 0.002070029266178608, + 0.0025742098223417997, + 0.004158603958785534, + 0.003504305612295866, + 0.005862615071237087, + 0.003434635465964675, + 0.0030495866667479277, + 0.004756602458655834, + 0.0048222532495856285, + 0.004440027289092541, + 0.007981069386005402, + 0.006774289067834616, + 0.01977398619055748, + 0.007123919669538736, + 0.005373829510062933, + 0.01108782272785902, + 0.008441154845058918, + 0.01080250646919012, + 0.0016487747197970748, + 0.0026228390634059906, + 0.0025031231343746185, + 0.003921521827578545, + 0.0014566100435331464, + 0.0020351035054773092, + 0.002586447400972247, + 0.0018612664425745606, + 0.004326526075601578, + 0.005465661641210318, + 0.004241368267685175, + 0.003114449791610241, + 0.004221324808895588, + 0.007248270325362682, + 0.003038718132302165, + 0.0050475201569497585, + 0.002618300961330533, + 0.005453027319163084, + 0.0021623738575726748, + 0.0029318274464458227, + 0.0039028965402394533, + 0.002486628480255604, + 0.002063833409920335, + 0.0023789163678884506, + 0.0035375196021050215, + 0.003879427444189787, + 0.0040883407928049564, + 0.0022374435793608427, + 0.003662914503365755, + 0.0044899918138980865, + 0.002096770331263542, + 0.0028095103334635496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.06207595765590668 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197180047631264 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.014197180047631264, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197180047631264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22822153568267822, + 0.13356627523899078, + 0.36265328526496887, + 0.06627847999334335, + 0.06241638585925102, + 0.05048155039548874, + 0.02327495440840721, + 0.022987553849816322, + 0.033860333263874054, + 0.01625967025756836 + ], + "credit_map": [ + 0.019069179892539978, + 0.018188653513789177, + 0.04535917937755585, + 0.038579925894737244, + 0.021604858338832855, + 0.030251527205109596, + 0.02385758049786091, + 0.031310610473155975, + 0.01867511309683323, + 0.006886352319270372, + 0.011286797001957893, + 0.010587955825030804, + 0.007522917352616787, + 0.03502094745635986, + 0.007611205335706472, + 0.035974983125925064, + 0.03002791851758957, + 0.04217178747057915, + 0.03483116254210472, + 0.04747949168086052, + 0.07043620944023132, + 0.03490137681365013, + 0.06481093913316727, + 0.03799436613917351, + 0.0057914964854717255, + 0.004897446371614933, + 0.008603520691394806, + 0.011201935820281506, + 0.007182249333709478, + 0.005407078191637993, + 0.015707213431596756, + 0.007487535942345858, + 0.006273704580962658, + 0.006083087995648384, + 0.009807197377085686, + 0.006890499964356422, + 0.006633165292441845, + 0.010404922068119049, + 0.008259773254394531, + 0.008064032532274723, + 0.004404129460453987, + 0.0052852691151201725, + 0.009215317666530609, + 0.005984808783978224, + 0.005253140348941088, + 0.00698197353631258, + 0.007707215845584869, + 0.005649694241583347, + 0.0018179569160565734, + 0.004300582688301802, + 0.002579299733042717, + 0.0032555125653743744, + 0.003053131280466914, + 0.002559554297477007, + 0.002568403724581003, + 0.0031405133195221424, + 0.0029271638486534357, + 0.0022042624186724424, + 0.0037645301781594753, + 0.0028937736060470343, + 0.0023800560738891363, + 0.002919985679909587, + 0.0032487940043210983, + 0.00264898594468832, + 0.003286655992269516, + 0.007168704643845558, + 0.003909974824637175, + 0.0034205433912575245, + 0.005029621534049511, + 0.004037574399262667, + 0.0034149379935115576, + 0.0035923169925808907, + 0.002250532852485776, + 0.002701792400330305, + 0.002028138842433691, + 0.0020213613752275705, + 0.0016161532839760184, + 0.0027623160276561975, + 0.0012682895176112652, + 0.001611083745956421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.014197180047631264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100250080227852 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.004100250080227852, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100250080227852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1914263665676117, + 0.17072536051273346, + 0.2380857765674591, + 0.05912734568119049, + 0.03974822163581848, + 0.12322929501533508, + 0.02791113778948784, + 0.031876418739557266, + 0.08846105635166168, + 0.02940908446907997 + ], + "credit_map": [ + 0.022450881078839302, + 0.023528246209025383, + 0.03076658956706524, + 0.02722373977303505, + 0.02054992876946926, + 0.022916032001376152, + 0.021366706117987633, + 0.02262425608932972, + 0.020446276292204857, + 0.005295401904731989, + 0.014217229560017586, + 0.008651839569211006, + 0.00783711951225996, + 0.05532831326127052, + 0.00890205055475235, + 0.0500471293926239, + 0.022019395604729652, + 0.021023139357566833, + 0.030095163732767105, + 0.02589455619454384, + 0.03769780322909355, + 0.03151397779583931, + 0.03938084468245506, + 0.030460890382528305, + 0.008631649427115917, + 0.0044821337796747684, + 0.007998914457857609, + 0.005031529348343611, + 0.006655908189713955, + 0.009745984338223934, + 0.006960826460272074, + 0.009620399214327335, + 0.003969553858041763, + 0.004194146022200584, + 0.005746495444327593, + 0.0047865211963653564, + 0.004799938295036554, + 0.004407970700412989, + 0.006877955049276352, + 0.004965638741850853, + 0.011384823359549046, + 0.01590188406407833, + 0.029511835426092148, + 0.008054233156144619, + 0.013156941160559654, + 0.013216438703238964, + 0.014813863672316074, + 0.01718927174806595, + 0.0030581324826925993, + 0.002741285366937518, + 0.0046407259069383144, + 0.0033441309351474047, + 0.0031927055679261684, + 0.002934602089226246, + 0.003826440079137683, + 0.004173113498836756, + 0.004091483075171709, + 0.0032831488642841578, + 0.004823756404221058, + 0.0034246868453919888, + 0.0036549363285303116, + 0.005326719488948584, + 0.0036489085759967566, + 0.0036227779928594828, + 0.008150859735906124, + 0.02073746919631958, + 0.008902938105165958, + 0.012266567908227444, + 0.013216182589530945, + 0.009097468107938766, + 0.008703726343810558, + 0.007385849021375179, + 0.00301756733097136, + 0.005840036552399397, + 0.002000454580411315, + 0.003285593818873167, + 0.0038564554415643215, + 0.0054420907981693745, + 0.002366648055613041, + 0.003600235329940915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.004100250080227852 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100250080227852 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.004100250080227852, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100250080227852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18986551463603973, + 0.2429557591676712, + 0.21152405440807343, + 0.028706828132271767, + 0.07256991416215897, + 0.08785198628902435, + 0.029480431228876114, + 0.025955501943826675, + 0.08223860710859299, + 0.028851432725787163 + ], + "credit_map": [ + 0.020073939114809036, + 0.016954787075519562, + 0.039857253432273865, + 0.02372279204428196, + 0.017475808039307594, + 0.030873851850628853, + 0.01657075248658657, + 0.024336321279406548, + 0.028379002586007118, + 0.009675630368292332, + 0.014489599503576756, + 0.011707596480846405, + 0.010944264009594917, + 0.07728555798530579, + 0.012764682993292809, + 0.07770942896604538, + 0.0188179649412632, + 0.02041499689221382, + 0.023359671235084534, + 0.023265711963176727, + 0.03865472599864006, + 0.02294164150953293, + 0.03906041383743286, + 0.0250089168548584, + 0.0032929053995758295, + 0.0022851144894957542, + 0.003203177358955145, + 0.003149287076666951, + 0.002871711738407612, + 0.005858014803379774, + 0.003376228269189596, + 0.004670390393584967, + 0.006403689738363028, + 0.008405364118516445, + 0.009306884370744228, + 0.009185949340462685, + 0.008954312652349472, + 0.007184424437582493, + 0.014705116860568523, + 0.008424175903201103, + 0.007105415686964989, + 0.010138975456357002, + 0.015953242778778076, + 0.009903313592076302, + 0.010187055915594101, + 0.011928042396903038, + 0.012152299284934998, + 0.010483642108738422, + 0.0029361448250710964, + 0.003187701338902116, + 0.004094615578651428, + 0.003850275417789817, + 0.002696211449801922, + 0.004313397686928511, + 0.0035680250730365515, + 0.004834061488509178, + 0.0026693418622016907, + 0.0019913113210350275, + 0.0038296522106975317, + 0.0028381948359310627, + 0.002473260276019573, + 0.005673522595316172, + 0.0026702932082116604, + 0.003809927264228463, + 0.007973714731633663, + 0.017445048317313194, + 0.008615360595285892, + 0.010508394800126553, + 0.012296807020902634, + 0.008913270197808743, + 0.00797694269567728, + 0.008509076200425625, + 0.0035731778480112553, + 0.0033547645434737206, + 0.0039572948589921, + 0.002886171918362379, + 0.0029054279439151287, + 0.004829107318073511, + 0.0028439932502806187, + 0.004501493647694588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.004100250080227852 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures a perfect drop of shimmering golden maple syrup landing on a dark, matte, non-reflective surface. The droplet becomes a viscous, shimmering dome, its high-gloss surface reflecting a soft light, imbuing it with a warm, honey-hued glow. As the droplet slowly solidifies, a transformation begins almost imperceptibly; deep within its core, a tiny, star-shaped point of crystallization emerges. From this epicenter, a complex crystalline structure steadily and deliberately blooms outward, spreading like delicate frost flowers through the liquid gold. The once-smooth surface gives way to a complex network of geometric facets, light catching each new edge and causing it to shimmer. The process is a silent, mesmerizing ballet of molecular transformations, the viscous liquid meticulously reshaped, ultimately solidifying into a hard, translucent, multi-faceted amber gemstone that captures and refracts light in a dazzling prismatic display, creating an atmosphere of quiet, natural alchemy and focused wonder.", + "seed": 31065, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_0/index_8076_seed_31065_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1163294674595818, + "vlm_logprob": -0.1163294674595818, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.1163294674595818, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030945848673582077 + ], + "opposite_logprob_tokens": [ + -3.7809457778930664 + ], + "logprob_reward": -0.030945848673582077, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.030945848673582077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08813924342393875, + 0.1179351955652237, + 0.5880507826805115, + 0.02589522860944271, + 0.020345421507954597, + 0.03722689673304558, + 0.026511769741773605, + 0.030214723199605942, + 0.03381730243563652, + 0.03186345845460892 + ], + "credit_map": [ + 0.012003675103187561, + 0.006971186492592096, + 0.014479567296802998, + 0.009098107926547527, + 0.0058331904001533985, + 0.018024293705821037, + 0.008125991560518742, + 0.013603231869637966, + 0.0056291851215064526, + 0.005188785493373871, + 0.007730595767498016, + 0.00915518682450056, + 0.004755503498017788, + 0.04248308762907982, + 0.005598373245447874, + 0.03739447146654129, + 0.024451429024338722, + 0.013036549091339111, + 0.019528860226273537, + 0.010226313024759293, + 0.011407860554754734, + 0.25717008113861084, + 0.013218404725193977, + 0.2390112578868866, + 0.0037581182550638914, + 0.003001269418746233, + 0.0038770174141973257, + 0.0026899687945842743, + 0.0032572392374277115, + 0.0029054968617856503, + 0.003980701323598623, + 0.002425418235361576, + 0.0026071658357977867, + 0.001542451442219317, + 0.0018676288891583681, + 0.001547151361592114, + 0.0024053077213466167, + 0.004823241848498583, + 0.002574677811935544, + 0.002977796131744981, + 0.006524186581373215, + 0.0027201129123568535, + 0.004434031434357166, + 0.0029989255126565695, + 0.0038623036816716194, + 0.006775489076972008, + 0.003466458059847355, + 0.006445388775318861, + 0.004210157319903374, + 0.002682378748431802, + 0.0026577769313007593, + 0.0022813365794718266, + 0.005066131707280874, + 0.0039058139082044363, + 0.003157560247927904, + 0.0025506133679300547, + 0.0030235436279326677, + 0.0034890081733465195, + 0.003815768053755164, + 0.004687183536589146, + 0.002626914530992508, + 0.002587482798844576, + 0.005595784168690443, + 0.004389036912471056, + 0.0034274584613740444, + 0.004432568792253733, + 0.0029191288631409407, + 0.0035819527693092823, + 0.00551098957657814, + 0.003819430014118552, + 0.005183172412216663, + 0.004942600149661303, + 0.0038384729996323586, + 0.00313012907281518, + 0.004022605251520872, + 0.0032093296758830547, + 0.0042562237940728664, + 0.003921516705304384, + 0.005884439684450626, + 0.003600743133574724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.030945848673582077 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233445167541504 + ], + "opposite_logprob_tokens": [ + -0.8983445167541504 + ], + "logprob_reward": -0.5233445167541504, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233445167541504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.165083646774292, + 0.2637515366077423, + 0.30534476041793823, + 0.07096660882234573, + 0.06598488986492157, + 0.0248773992061615, + 0.030690962448716164, + 0.025617757812142372, + 0.03100469335913658, + 0.01667771115899086 + ], + "credit_map": [ + 0.01892833225429058, + 0.011650322005152702, + 0.03954532742500305, + 0.011159568093717098, + 0.010543334297835827, + 0.027079198509454727, + 0.010288603603839874, + 0.035888973623514175, + 0.02405337430536747, + 0.00994010828435421, + 0.032925840467214584, + 0.012156968005001545, + 0.009448261000216007, + 0.08453425019979477, + 0.00876743346452713, + 0.08192532509565353, + 0.03258325159549713, + 0.011024030856788158, + 0.029612980782985687, + 0.010076221078634262, + 0.009107602760195732, + 0.11001363396644592, + 0.011072029359638691, + 0.0918550193309784, + 0.010055738501250744, + 0.006920025683939457, + 0.01042243279516697, + 0.0052413311786949635, + 0.011532514356076717, + 0.01285095326602459, + 0.0074011110700666904, + 0.006542511284351349, + 0.008206730708479881, + 0.007252609357237816, + 0.006912729237228632, + 0.004718323238193989, + 0.01610933057963848, + 0.008542423136532307, + 0.008020338602364063, + 0.006222415715456009, + 0.005207524169236422, + 0.0017488346202298999, + 0.0027335293125361204, + 0.0022286793682724237, + 0.001929080463014543, + 0.004299689084291458, + 0.0034063186030834913, + 0.0033237473107874393, + 0.004440122749656439, + 0.00295727769844234, + 0.0035026532132178545, + 0.0025890886317938566, + 0.006440899334847927, + 0.004174703266471624, + 0.0036189488600939512, + 0.0029672712553292513, + 0.002907032612711191, + 0.002289221389219165, + 0.004853425081819296, + 0.0036627999506890774, + 0.0022372687235474586, + 0.004063130356371403, + 0.002631679642945528, + 0.00297320238314569, + 0.002713344991207123, + 0.0037320980336517096, + 0.0028804324101656675, + 0.004205095581710339, + 0.006350757088512182, + 0.0035397398751229048, + 0.0050415657460689545, + 0.002541664522141218, + 0.0020125573500990868, + 0.0017725932411849499, + 0.001977081410586834, + 0.0023331406991928816, + 0.0013835130957886577, + 0.0018597525777295232, + 0.0034996161703020334, + 0.0018394575454294682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.5233445167541504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058335766196251 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058335766196251, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058335766196251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20910580456256866, + 0.14975449442863464, + 0.17325353622436523, + 0.10105302184820175, + 0.06270993500947952, + 0.11265528202056885, + 0.046323224902153015, + 0.028413500636816025, + 0.04322981834411621, + 0.07350144535303116 + ], + "credit_map": [ + 0.040561821311712265, + 0.013546749949455261, + 0.02766663208603859, + 0.013113805092871189, + 0.013544881716370583, + 0.05317772179841995, + 0.010999697260558605, + 0.036494504660367966, + 0.008763054385781288, + 0.009251170791685581, + 0.012919537723064423, + 0.00870567373931408, + 0.010440039448440075, + 0.04967299476265907, + 0.005925674457103014, + 0.0440763458609581, + 0.008994483388960361, + 0.006242855452001095, + 0.007852379232645035, + 0.00545016722753644, + 0.008831199258565903, + 0.06642041355371475, + 0.0055658710189163685, + 0.06389616429805756, + 0.013593257404863834, + 0.009751088917255402, + 0.009896700270473957, + 0.009213375858962536, + 0.015119184739887714, + 0.017967548221349716, + 0.016350138932466507, + 0.009161731228232384, + 0.008140256628394127, + 0.005276038311421871, + 0.005538354627788067, + 0.003854910144582391, + 0.007988182827830315, + 0.0156676284968853, + 0.006567959673702717, + 0.00967660453170538, + 0.012648706324398518, + 0.009312236681580544, + 0.014812254346907139, + 0.011286710388958454, + 0.01872641034424305, + 0.01651862822473049, + 0.012989590875804424, + 0.016360746696591377, + 0.0053292508237063885, + 0.00231469189748168, + 0.0035522368270903826, + 0.002570678712800145, + 0.005440485198050737, + 0.013468164019286633, + 0.0033363986294716597, + 0.010311315767467022, + 0.0030653728172183037, + 0.003193352837115526, + 0.004127830266952515, + 0.004845285788178444, + 0.002466283505782485, + 0.002905159955844283, + 0.0034431335516273975, + 0.004367080517113209, + 0.00469738757237792, + 0.004365874454379082, + 0.00292955175973475, + 0.006907264702022076, + 0.004423509817570448, + 0.005934915971010923, + 0.006911769043654203, + 0.007059543859213591, + 0.008153567090630531, + 0.009805097244679928, + 0.005194338038563728, + 0.008618897758424282, + 0.01103546004742384, + 0.008480004034936428, + 0.013812076300382614, + 0.008402002975344658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.02058335766196251 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258182529360056 + ], + "opposite_logprob_tokens": [ + -5.255258083343506 + ], + "logprob_reward": -0.005258182529360056, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258182529360056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06876084208488464, + 0.1309884637594223, + 0.4626983106136322, + 0.0448475182056427, + 0.04641271010041237, + 0.054501961916685104, + 0.022661378607153893, + 0.015043864957988262, + 0.054382652044296265, + 0.09970234334468842 + ], + "credit_map": [ + 0.007474518846720457, + 0.009552947245538235, + 0.010224186815321445, + 0.007907901890575886, + 0.007470271550118923, + 0.009341935627162457, + 0.006402556784451008, + 0.010386521928012371, + 0.006098564248532057, + 0.004847490228712559, + 0.008419911377131939, + 0.005945471581071615, + 0.005971034988760948, + 0.05013970285654068, + 0.0038021658547222614, + 0.045764125883579254, + 0.018502408638596535, + 0.008617092855274677, + 0.012871515937149525, + 0.006410941015928984, + 0.007079523988068104, + 0.20800155401229858, + 0.009368719533085823, + 0.19184653460979462, + 0.006135035306215286, + 0.005707878619432449, + 0.004594107624143362, + 0.004678546451032162, + 0.007971893064677715, + 0.005657928995788097, + 0.006223222240805626, + 0.0038789017125964165, + 0.0061631216667592525, + 0.0037382221780717373, + 0.005965552758425474, + 0.0033367748837918043, + 0.0062536559998989105, + 0.008737350814044476, + 0.006505197379738092, + 0.005712834652513266, + 0.006364660803228617, + 0.0036026614252477884, + 0.007186316419392824, + 0.005903370678424835, + 0.006352399010211229, + 0.007990557700395584, + 0.008725186809897423, + 0.008376812562346458, + 0.0030482669826596975, + 0.002187273232266307, + 0.002426900900900364, + 0.002199650974944234, + 0.00400560861453414, + 0.0029484042897820473, + 0.002959742909297347, + 0.0028855304699391127, + 0.0015074763214215636, + 0.0025464303325861692, + 0.0017471733735874295, + 0.002307840855792165, + 0.0012824023142457008, + 0.0017030549934133887, + 0.002183312550187111, + 0.0017661751480773091, + 0.005267734173685312, + 0.008095139637589455, + 0.005044073332101107, + 0.009478594176471233, + 0.007570808753371239, + 0.004680428188294172, + 0.008518307469785213, + 0.005727570969611406, + 0.010328853502869606, + 0.009954451583325863, + 0.007382811978459358, + 0.013048101216554642, + 0.0115634361281991, + 0.012997548095881939, + 0.022985272109508514, + 0.011441871523857117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.005258182529360056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001515431678853929 + ], + "opposite_logprob_tokens": [ + -6.5015153884887695 + ], + "logprob_reward": -0.001515431678853929, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001515431678853929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13624536991119385, + 0.12413066625595093, + 0.28447791934013367, + 0.031081626191735268, + 0.06757020205259323, + 0.06571254134178162, + 0.02014588750898838, + 0.021082695573568344, + 0.155096173286438, + 0.0944570004940033 + ], + "credit_map": [ + 0.02278275042772293, + 0.007453409489244223, + 0.021181587129831314, + 0.01122195366770029, + 0.009814857505261898, + 0.03092844784259796, + 0.010120992548763752, + 0.02274136431515217, + 0.010777287185192108, + 0.006138312630355358, + 0.009185953065752983, + 0.00514365267008543, + 0.006474638823419809, + 0.042792920023202896, + 0.003781776875257492, + 0.03983611240983009, + 0.012591780163347721, + 0.005131135229021311, + 0.007971260696649551, + 0.003801463171839714, + 0.0055869328789412975, + 0.12837906181812286, + 0.005361475050449371, + 0.11565481871366501, + 0.00591515377163887, + 0.002685497747734189, + 0.0045909336768090725, + 0.002897307276725769, + 0.0032326863147318363, + 0.004989763256162405, + 0.003616063855588436, + 0.003154219128191471, + 0.010778888128697872, + 0.007700733374804258, + 0.008969438262283802, + 0.006644025910645723, + 0.011704199016094208, + 0.008301710709929466, + 0.0074934945441782475, + 0.00597771629691124, + 0.007945599965751171, + 0.004952609073370695, + 0.008158063516020775, + 0.007275602780282497, + 0.010260271839797497, + 0.009382632561028004, + 0.008844355121254921, + 0.008893407881259918, + 0.0022563624661415815, + 0.001410350319929421, + 0.0044282227754592896, + 0.002308927709236741, + 0.0026177179533988237, + 0.0021571831312030554, + 0.0030738976784050465, + 0.0018932267557829618, + 0.002708274172618985, + 0.0024165003560483456, + 0.0026963294949382544, + 0.003024044446647167, + 0.0020132341887801886, + 0.002838814863935113, + 0.0028248403687030077, + 0.0025606579147279263, + 0.01563369482755661, + 0.016369950026273727, + 0.013851874507963657, + 0.029296325519680977, + 0.02063112147152424, + 0.018268190324306488, + 0.02411816455423832, + 0.016926851123571396, + 0.010413303971290588, + 0.010617639869451523, + 0.006446192506700754, + 0.010874979197978973, + 0.012133997865021229, + 0.01272357627749443, + 0.019993405789136887, + 0.011253907345235348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.001515431678853929 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up captures a perfect drop of shimmering golden maple syrup landing on a dark, matte, non-reflective surface. The droplet becomes a viscous, shimmering dome, its high-gloss surface reflecting a soft light, imbuing it with a warm, honey-hued glow. As the droplet slowly solidifies, a transformation begins almost imperceptibly; deep within its core, a tiny, star-shaped point of crystallization emerges. From this epicenter, a complex crystalline structure steadily and deliberately blooms outward, spreading like delicate frost flowers through the liquid gold. The once-smooth surface gives way to a complex network of geometric facets, light catching each new edge and causing it to shimmer. The process is a silent, mesmerizing ballet of molecular transformations, the viscous liquid meticulously reshaped, ultimately solidifying into a hard, translucent, multi-faceted amber gemstone that captures and refracts light in a dazzling prismatic display, creating an atmosphere of quiet, natural alchemy and focused wonder.", + "seed": 31066, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_0/index_8076_seed_31066_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0945547058712691, + "vlm_logprob": -0.0945547058712691, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.0945547058712691, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024174215272068977 + ], + "opposite_logprob_tokens": [ + -4.024174213409424 + ], + "logprob_reward": -0.024174215272068977, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024174215272068977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08892325311899185, + 0.08904709666967392, + 0.08491740375757217, + 0.022159120067954063, + 0.024802932515740395, + 0.125051349401474, + 0.050966642796993256, + 0.07081107795238495, + 0.028448693454265594, + 0.4148724675178528 + ], + "credit_map": [ + 0.010023688897490501, + 0.010141586884856224, + 0.008330540731549263, + 0.013211579993367195, + 0.013549050316214561, + 0.009433962404727936, + 0.013366332277655602, + 0.010866502299904823, + 0.013039061799645424, + 0.012218098156154156, + 0.014630860649049282, + 0.007044808939099312, + 0.010460791178047657, + 0.009703025221824646, + 0.009418750181794167, + 0.012531689368188381, + 0.018189679831266403, + 0.004251874051988125, + 0.012399966828525066, + 0.0027144562918692827, + 0.003175250720232725, + 0.013307467103004456, + 0.006283854134380817, + 0.024594847112894058, + 0.004214689135551453, + 0.0025067413225769997, + 0.002391856862232089, + 0.0021880008280277252, + 0.0032675988040864468, + 0.002431026194244623, + 0.0019864945206791162, + 0.0031727110035717487, + 0.0037121169734746218, + 0.0034117880277335644, + 0.002547287615016103, + 0.0019917646422982216, + 0.00498048635199666, + 0.0031220733653753996, + 0.0025724333245307207, + 0.002464980585500598, + 0.012953048571944237, + 0.013555886223912239, + 0.01302450243383646, + 0.012738281860947609, + 0.026470158249139786, + 0.015634683892130852, + 0.013908123597502708, + 0.016766661778092384, + 0.0069115343503654, + 0.005276271607726812, + 0.004924661945551634, + 0.006456373725086451, + 0.008432933129370213, + 0.006810963153839111, + 0.006170614622533321, + 0.005983283743262291, + 0.006962902843952179, + 0.009827537462115288, + 0.009502201341092587, + 0.006953968200832605, + 0.01037506852298975, + 0.008357655256986618, + 0.00592407863587141, + 0.012907656840980053, + 0.0034109519328922033, + 0.005712714046239853, + 0.0018209333065897226, + 0.0031560612842440605, + 0.00447429483756423, + 0.003359791822731495, + 0.0035693603567779064, + 0.0029445854015648365, + 0.013202369213104248, + 0.009697862900793552, + 0.005330859217792749, + 0.008161026984453201, + 0.17811492085456848, + 0.18592412769794464, + 0.006742105353623629, + 0.007699161767959595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.024174215272068977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289219379425049 + ], + "opposite_logprob_tokens": [ + -1.0539219379425049 + ], + "logprob_reward": -0.4289219379425049, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289219379425049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18347999453544617, + 0.20207874476909637, + 0.08163755387067795, + 0.03109581582248211, + 0.02451765537261963, + 0.2843289077281952, + 0.024816153571009636, + 0.020836496725678444, + 0.03989546000957489, + 0.10731329768896103 + ], + "credit_map": [ + 0.020864980295300484, + 0.026528839021921158, + 0.022721458226442337, + 0.01690879836678505, + 0.017171146348118782, + 0.021788369864225388, + 0.03150063008069992, + 0.025995755568146706, + 0.027236156165599823, + 0.013897867873311043, + 0.03174934163689613, + 0.0103386789560318, + 0.014784947037696838, + 0.030672037973999977, + 0.013401840813457966, + 0.05999785289168358, + 0.013341307640075684, + 0.006624337751418352, + 0.009363992139697075, + 0.0033799228258430958, + 0.00402802461758256, + 0.013926587998867035, + 0.006335738115012646, + 0.024637633934617043, + 0.0054052844643592834, + 0.003217128338292241, + 0.0047738575376570225, + 0.0030549466609954834, + 0.004307132214307785, + 0.003555465256795287, + 0.0029467896092683077, + 0.0038352073170244694, + 0.0028757029213011265, + 0.004561577457934618, + 0.0026139146648347378, + 0.0018756535137072206, + 0.004608902148902416, + 0.0030058356933295727, + 0.0022031653206795454, + 0.0027729046996682882, + 0.02871014177799225, + 0.032136913388967514, + 0.030171483755111694, + 0.03481324017047882, + 0.06402449309825897, + 0.03787899389863014, + 0.031164182350039482, + 0.02542944625020027, + 0.003116121981292963, + 0.0035588578321039677, + 0.0027642250061035156, + 0.002851589350029826, + 0.003647048259153962, + 0.0031964764930307865, + 0.0027530204970389605, + 0.002928812289610505, + 0.00298287277109921, + 0.0027607169467955828, + 0.0021694155875593424, + 0.0023038561921566725, + 0.003146205563098192, + 0.0023706196807324886, + 0.002136804163455963, + 0.0029660030268132687, + 0.00480244355276227, + 0.004837973043322563, + 0.003675994463264942, + 0.005284181330353022, + 0.003957626409828663, + 0.004556188825517893, + 0.0064164623618125916, + 0.006364586763083935, + 0.0033877440728247166, + 0.0037110215052962303, + 0.001761321211233735, + 0.0033932975493371487, + 0.0446968711912632, + 0.04610026255249977, + 0.0021364663261920214, + 0.0021263090893626213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.4289219379425049 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542534619569778 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542534619569778, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542534619569778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11472053080797195, + 0.11093010753393173, + 0.04425157606601715, + 0.043871112167835236, + 0.03398137167096138, + 0.4395860433578491, + 0.03802384436130524, + 0.06994003057479858, + 0.038690000772476196, + 0.0660053938627243 + ], + "credit_map": [ + 0.010719148442149162, + 0.016944192349910736, + 0.012480942532420158, + 0.009644872508943081, + 0.013459700159728527, + 0.013558805920183659, + 0.020126188173890114, + 0.01778668910264969, + 0.019929833710193634, + 0.009900187142193317, + 0.023873526602983475, + 0.008746977895498276, + 0.009757776744663715, + 0.012304245494306087, + 0.008848393335938454, + 0.0175691619515419, + 0.007685334421694279, + 0.00562054431065917, + 0.005727177485823631, + 0.004419115837663412, + 0.004160180222243071, + 0.004431582055985928, + 0.005954412743449211, + 0.0062532261945307255, + 0.007338273338973522, + 0.0055450694635510445, + 0.005197996273636818, + 0.004622775595635176, + 0.005472766701132059, + 0.004465950652956963, + 0.0061928569339215755, + 0.005035424139350653, + 0.00579413864761591, + 0.004745890852063894, + 0.004696747753769159, + 0.0033894397784024477, + 0.004524471703916788, + 0.0037129514385014772, + 0.003066505305469036, + 0.00405122758820653, + 0.04497634992003441, + 0.05039370432496071, + 0.04511018097400665, + 0.0514545775949955, + 0.09956984221935272, + 0.05991865694522858, + 0.04783264920115471, + 0.04033010080456734, + 0.004011156503111124, + 0.0038398371543735266, + 0.005504264496266842, + 0.0076584224589169025, + 0.004169483203440905, + 0.003923971671611071, + 0.0042742593213915825, + 0.004642452113330364, + 0.0070197745226323605, + 0.010626778937876225, + 0.009409249760210514, + 0.007016678340733051, + 0.01007752027362585, + 0.00863677728921175, + 0.005791214760392904, + 0.011362033896148205, + 0.004327962175011635, + 0.004278352949768305, + 0.0039871856570243835, + 0.0045387097634375095, + 0.005848462227731943, + 0.004931618459522724, + 0.005780857987701893, + 0.004996851086616516, + 0.0023325264919549227, + 0.002920240629464388, + 0.0014329022960737348, + 0.0018615220906212926, + 0.026223400607705116, + 0.027552299201488495, + 0.0016676292289048433, + 0.002014865167438984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.012542534619569778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642304033041 + ], + "opposite_logprob_tokens": [ + -5.379642486572266 + ], + "logprob_reward": -0.004642304033041, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642304033041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11414634436368942, + 0.061637427657842636, + 0.059531085193157196, + 0.04405352100729942, + 0.04136473685503006, + 0.26139116287231445, + 0.04320855066180229, + 0.06141132116317749, + 0.10733414441347122, + 0.20592166483402252 + ], + "credit_map": [ + 0.012755821458995342, + 0.022159980610013008, + 0.013707759790122509, + 0.009640008211135864, + 0.016366107389330864, + 0.012095071375370026, + 0.015727635473012924, + 0.011693977750837803, + 0.00809468887746334, + 0.010138066485524178, + 0.007539810612797737, + 0.0066557941026985645, + 0.008048644289374352, + 0.006462777964770794, + 0.00608908711001277, + 0.0086085619404912, + 0.009897115640342236, + 0.006915910169482231, + 0.007715578190982342, + 0.004524352960288525, + 0.007425753399729729, + 0.006932518910616636, + 0.006424910854548216, + 0.009694946929812431, + 0.007278212811797857, + 0.0050131212919950485, + 0.005016855429857969, + 0.005222786217927933, + 0.005286193918436766, + 0.004316865000873804, + 0.0069666896015405655, + 0.004952800460159779, + 0.006478188093751669, + 0.008201908320188522, + 0.004089417867362499, + 0.004287515301257372, + 0.006443215534090996, + 0.004869952332228422, + 0.0029064337722957134, + 0.004088107496500015, + 0.026860622689127922, + 0.02870086207985878, + 0.03056841902434826, + 0.029397400096058846, + 0.05832008272409439, + 0.03538523241877556, + 0.027376679703593254, + 0.024781854823231697, + 0.00493517704308033, + 0.003864691127091646, + 0.004687784239649773, + 0.006764626130461693, + 0.00574036780744791, + 0.004276872146874666, + 0.005427340045571327, + 0.007511691190302372, + 0.006093260366469622, + 0.012264556251466274, + 0.004658237565308809, + 0.006778158247470856, + 0.01081330981105566, + 0.008610161952674389, + 0.006606705021113157, + 0.005586939863860607, + 0.01462995819747448, + 0.012246270664036274, + 0.01008602138608694, + 0.010194951668381691, + 0.016886157914996147, + 0.012883543036878109, + 0.013564135879278183, + 0.01684311218559742, + 0.0065036341547966, + 0.0068586524575948715, + 0.004051533993333578, + 0.005850344896316528, + 0.08526360243558884, + 0.08619187772274017, + 0.0046767969615757465, + 0.006525249220430851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.004642304033041 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a drop of golden maple syrup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the syrup land on a dark, matte, non-reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a star-shaped point of crystallization emerge from the core of the droplet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the liquid transform into a hard, translucent, multi-faceted amber gemstone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final gemstone display a dazzling prismatic light refraction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024925374891608953 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024925374891608953, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024925374891608953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0726068988442421, + 0.08047574013471603, + 0.059041500091552734, + 0.024730876088142395, + 0.03129720315337181, + 0.0618617981672287, + 0.0699184462428093, + 0.04181033372879028, + 0.042048338800668716, + 0.5162088871002197 + ], + "credit_map": [ + 0.007848881185054779, + 0.009537514299154282, + 0.010949010029435158, + 0.006701625883579254, + 0.009142148308455944, + 0.008632168173789978, + 0.00973941944539547, + 0.010056128725409508, + 0.009679563343524933, + 0.012778326869010925, + 0.011066640727221966, + 0.0063444096595048904, + 0.010837176814675331, + 0.008860712870955467, + 0.008705570362508297, + 0.0122033366933465, + 0.009931476786732674, + 0.005696375388652086, + 0.007288343273103237, + 0.0031533907167613506, + 0.004047303460538387, + 0.008386055007576942, + 0.005020183511078358, + 0.015518372878432274, + 0.003120937617495656, + 0.003617403795942664, + 0.002485064323991537, + 0.002539047971367836, + 0.0036423855926841497, + 0.002668391913175583, + 0.0028466901276260614, + 0.0038109556771814823, + 0.005268784239888191, + 0.0038640638813376427, + 0.0038540016394108534, + 0.004176181275397539, + 0.004489274695515633, + 0.003413795493543148, + 0.0027900084387511015, + 0.0034410939551889896, + 0.009786291979253292, + 0.00706374179571867, + 0.01057847123593092, + 0.006293479818850756, + 0.006158808246254921, + 0.005975937005132437, + 0.0047658272087574005, + 0.011239240877330303, + 0.00715789757668972, + 0.008499990217387676, + 0.006869047414511442, + 0.012487108819186687, + 0.010012735612690449, + 0.008172555826604366, + 0.008640112355351448, + 0.00807899422943592, + 0.004531234037131071, + 0.006466270890086889, + 0.0034788469783961773, + 0.004344074986875057, + 0.007473581004887819, + 0.005228585097938776, + 0.004980041645467281, + 0.005307698156684637, + 0.0038967474829405546, + 0.00548033369705081, + 0.004612076096236706, + 0.0052698152139782906, + 0.0069632683880627155, + 0.005746722687035799, + 0.005746791604906321, + 0.004332581534981728, + 0.01582738198339939, + 0.01227277796715498, + 0.007061911281198263, + 0.00945138931274414, + 0.22238455712795258, + 0.23170700669288635, + 0.007946175523102283, + 0.009557662531733513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0024925374891608953 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_11/index_tensor([8289, 8289])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_11/index_tensor([8289, 8289])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..23bf2f3fc746e14d16d067d349a36ce38d196ed4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_11/index_tensor([8289, 8289])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a slowly advancing medium shot, captured from a slight upward angle, we are transported deep into a misty Eastern wonderland. A gigantic leaf-tailed gecko rests peacefully on a vine entwined with an ancient tree, emitting a soft blue-green glow. Its skin, a complex camouflage pattern reminiscent of millennia-old bark, is covered in mottled dark green and earthy brown mosses, punctuated by patches of silvery-gray lichen, almost blending seamlessly into the background. Its enormous, luminous amber eyes, with vertical pupils, gleam with a glimmer of all-seeing wisdom. Its signature broad, leaf-like tail, shaped like a withered autumn leaf, with naturally broken edges and distinct veins, trembles ever so slightly, almost imperceptibly, in sync with the invisible aura flowing around it. A thick, milky fog slowly drifts through the forest. The vine's bioluminescence illuminates the surrounding mist and the shimmering particles of light floating in the air, casting shifting shadows on the gecko's body. The entire picture is presented in a cinematic 16:9 aspect ratio and 4K high-definition resolution, with every detail clearly visible, creating an atmosphere of extreme tranquility, solemnity and ancient mystery.", + "seed": 9132, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_11/index_8289_seed_9132_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03392167054116726, + "vlm_logprob": -0.03392167054116726, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.03392167054116726, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016801515594124794 + ], + "opposite_logprob_tokens": [ + -4.141801357269287 + ], + "logprob_reward": -0.016801515594124794, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016801515594124794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09635978937149048, + 0.10504206269979477, + 0.020160486921668053, + 0.056834325194358826, + 0.03804219886660576, + 0.02072956971824169, + 0.6255595088005066, + 0.017048465088009834, + 0.0070166210643947124, + 0.013207007199525833 + ], + "credit_map": [ + 0.01704975962638855, + 0.0067877802066504955, + 0.009093983098864555, + 0.006716529838740826, + 0.007415803149342537, + 0.011710356920957565, + 0.013546477071940899, + 0.024039097130298615, + 0.02296021394431591, + 0.008951202966272831, + 0.01158803142607212, + 0.00675000948831439, + 0.009532525204122066, + 0.01264183409512043, + 0.0111794238910079, + 0.02143882028758526, + 0.002303518820554018, + 0.0022436855360865593, + 0.0020957719534635544, + 0.0016835960559546947, + 0.0015914220130071044, + 0.0035524542909115553, + 0.00392456678673625, + 0.002765470650047064, + 0.012696320191025734, + 0.0027866081800311804, + 0.0023051821626722813, + 0.002838890301063657, + 0.002363006118685007, + 0.004321909975260496, + 0.020684989169239998, + 0.008837416768074036, + 0.009060882963240147, + 0.0019890356343239546, + 0.0024836992379277945, + 0.0022442180197685957, + 0.0020562715362757444, + 0.0031429631635546684, + 0.011309850960969925, + 0.0057552773505449295, + 0.002242528833448887, + 0.002108983462676406, + 0.002874762285500765, + 0.0014666010392829776, + 0.002192700281739235, + 0.00292846467345953, + 0.0031132472213357687, + 0.0038022813387215137, + 0.11128506064414978, + 0.02776995860040188, + 0.02222399227321148, + 0.028581367805600166, + 0.030472854152321815, + 0.0181833915412426, + 0.24975204467773438, + 0.1372908502817154, + 0.0024156211875379086, + 0.002514455234631896, + 0.0012977502774447203, + 0.0021537630818784237, + 0.002377250464633107, + 0.001700921100564301, + 0.0020618620328605175, + 0.002526842290535569, + 0.0009203276713378727, + 0.0009802627610042691, + 0.0006389064365066588, + 0.0007548296125605702, + 0.000985070364549756, + 0.0008851967286318541, + 0.0008632839890196919, + 0.0009887428022921085, + 0.0019511774880811572, + 0.000883706146851182, + 0.0007528018904849887, + 0.0007264241576194763, + 0.001064399373717606, + 0.000981836230494082, + 0.004389577079564333, + 0.0024570843670517206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.016801515594124794 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302631452679634 + ], + "opposite_logprob_tokens": [ + -3.7733025550842285 + ], + "logprob_reward": -0.023302631452679634, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302631452679634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1615845412015915, + 0.07632351666688919, + 0.012914352118968964, + 0.02962162345647812, + 0.037345465272665024, + 0.026132691651582718, + 0.5298436284065247, + 0.04352924972772598, + 0.022679848596453667, + 0.06002508103847504 + ], + "credit_map": [ + 0.04280813783407211, + 0.007229492533951998, + 0.005678717978298664, + 0.005385121330618858, + 0.008494626730680466, + 0.008710354566574097, + 0.026380300521850586, + 0.05689779296517372, + 0.01575302518904209, + 0.007940584793686867, + 0.0076933992095291615, + 0.004986062180250883, + 0.008047491312026978, + 0.008394816890358925, + 0.008338646963238716, + 0.015169491060078144, + 0.0017157564871013165, + 0.0016754643293097615, + 0.0012636393075808883, + 0.0014470084570348263, + 0.0017783364746719599, + 0.001657667919062078, + 0.001612896448932588, + 0.0017635822296142578, + 0.005227501038461924, + 0.0016908823745325208, + 0.0012168639805167913, + 0.0016505579696968198, + 0.0012651047436520457, + 0.0012778728269040585, + 0.011376854032278061, + 0.005915985908359289, + 0.007409247104078531, + 0.002110769273713231, + 0.002425875049084425, + 0.002148340456187725, + 0.001748288981616497, + 0.0019803838804364204, + 0.013468315824866295, + 0.006054244935512543, + 0.0031766260508447886, + 0.0025853216648101807, + 0.004136068280786276, + 0.002194530563428998, + 0.003435167483985424, + 0.0032083257101476192, + 0.0036834627389907837, + 0.003713187761604786, + 0.09355932474136353, + 0.024147043004631996, + 0.019157150760293007, + 0.024425353854894638, + 0.02696315385401249, + 0.015996357426047325, + 0.2101622223854065, + 0.11543300002813339, + 0.0064692203886806965, + 0.004997322801500559, + 0.0030761538073420525, + 0.004430137574672699, + 0.00582353537902236, + 0.003162126522511244, + 0.009678095579147339, + 0.005892657209187746, + 0.0028151001315563917, + 0.0015603643842041492, + 0.0026716270949691534, + 0.0014782121870666742, + 0.001420823740772903, + 0.00248486315831542, + 0.0063149998895823956, + 0.003933859057724476, + 0.008681165054440498, + 0.004267598036676645, + 0.004635854158550501, + 0.003013042965903878, + 0.004160176496952772, + 0.005431569181382656, + 0.01885436661541462, + 0.01098131574690342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.023302631452679634 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540886178612709 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540886178612709, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540886178612709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11598310619592667, + 0.05914410948753357, + 0.021263405680656433, + 0.09025942534208298, + 0.04690210893750191, + 0.04164204001426697, + 0.5171788930892944, + 0.03434957563877106, + 0.015939893200993538, + 0.05733746290206909 + ], + "credit_map": [ + 0.025332322344183922, + 0.007141562644392252, + 0.004527619108557701, + 0.005279132630676031, + 0.008115381002426147, + 0.008984689600765705, + 0.020077334716916084, + 0.03652506321668625, + 0.012727880850434303, + 0.006198266055434942, + 0.0057777706533670425, + 0.004145494196563959, + 0.007440612185746431, + 0.006615332793444395, + 0.0058926925994455814, + 0.010346060618758202, + 0.0031266247387975454, + 0.0025699713733047247, + 0.002523059956729412, + 0.0017890341114252806, + 0.0020452316384762526, + 0.0021604723297059536, + 0.003022082382813096, + 0.0040269289165735245, + 0.021490616723895073, + 0.0036784163676202297, + 0.0040025352500379086, + 0.003999476786702871, + 0.00334092415869236, + 0.003349612932652235, + 0.03585037216544151, + 0.014547468163073063, + 0.010978206060826778, + 0.002737615257501602, + 0.0030906079337000847, + 0.002565104980021715, + 0.002818032866343856, + 0.004617531783878803, + 0.012847495265305042, + 0.007247516419738531, + 0.004708180669695139, + 0.0023257210850715637, + 0.0026657457929104567, + 0.002656956436112523, + 0.0026450997684150934, + 0.003986352123320103, + 0.014092954806983471, + 0.008561030961573124, + 0.0892406553030014, + 0.024999748915433884, + 0.019276456907391548, + 0.024362238124012947, + 0.030582671985030174, + 0.016854582354426384, + 0.20089371502399445, + 0.11096880584955215, + 0.004401586018502712, + 0.005137384869158268, + 0.003656485117971897, + 0.003206827910616994, + 0.0048542749136686325, + 0.00351635180413723, + 0.004648271016776562, + 0.004928393755108118, + 0.001781321712769568, + 0.002146112732589245, + 0.0015224595554172993, + 0.0016656594816595316, + 0.0025487097445875406, + 0.0022872593253850937, + 0.0018784565618261695, + 0.002109913621097803, + 0.009520864114165306, + 0.0032136919908225536, + 0.0031839015427976847, + 0.0022988691926002502, + 0.002983327256515622, + 0.0033624975476413965, + 0.02109859138727188, + 0.011675721034407616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.012540886178612709 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203136593103409 + ], + "opposite_logprob_tokens": [ + -2.8120312690734863 + ], + "logprob_reward": -0.06203136593103409, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203136593103409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04750416800379753, + 0.042598970234394073, + 0.01997012086212635, + 0.1792747527360916, + 0.03857659548521042, + 0.052707891911268234, + 0.5319716930389404, + 0.048777103424072266, + 0.017683200538158417, + 0.02093549817800522 + ], + "credit_map": [ + 0.006975009571760893, + 0.0073408521711826324, + 0.005452147219330072, + 0.004296006169170141, + 0.007111497689038515, + 0.006320865824818611, + 0.0052008191123604774, + 0.004806974902749062, + 0.009184750728309155, + 0.004362355452030897, + 0.004101634491235018, + 0.003267466789111495, + 0.0034892053809016943, + 0.004763713572174311, + 0.004815550521016121, + 0.008614301681518555, + 0.003021552227437496, + 0.0039509995840489864, + 0.0018167837988585234, + 0.0021876029204577208, + 0.0025838776491582394, + 0.0017706144135445356, + 0.002215856686234474, + 0.0024228361435234547, + 0.04039735719561577, + 0.006999558303505182, + 0.00668670330196619, + 0.008034361526370049, + 0.006763112731277943, + 0.0052392794750630856, + 0.07360514253377914, + 0.03154924884438515, + 0.008741283789277077, + 0.003665936877951026, + 0.0022434901911765337, + 0.002626837696880102, + 0.0029237563721835613, + 0.004724577534943819, + 0.007929475046694279, + 0.005721245892345905, + 0.007104632910341024, + 0.004969725385308266, + 0.003591385670006275, + 0.0034328089095652103, + 0.007390204351395369, + 0.006308342330157757, + 0.013417325913906097, + 0.006493472959846258, + 0.09204722940921783, + 0.0238370168954134, + 0.01835518330335617, + 0.02500159665942192, + 0.026199929416179657, + 0.01600145921111107, + 0.21278910338878632, + 0.11774017661809921, + 0.0058819628320634365, + 0.005255891941487789, + 0.0029406326357275248, + 0.0030643038917332888, + 0.008167602121829987, + 0.00314136128872633, + 0.01238085888326168, + 0.007944494485855103, + 0.002130189910531044, + 0.0027844305150210857, + 0.0017035453347489238, + 0.002054411917924881, + 0.002727660583332181, + 0.0019631064496934414, + 0.0022743830922991037, + 0.0020454730838537216, + 0.0027739237993955612, + 0.0029006986878812313, + 0.0013915225863456726, + 0.001589597319252789, + 0.002076916629448533, + 0.0036550345830619335, + 0.003608835395425558, + 0.0029389704577624798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.06203136593103409 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493195354938507 + ], + "opposite_logprob_tokens": [ + -2.929931879043579 + ], + "logprob_reward": -0.05493195354938507, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493195354938507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061772577464580536, + 0.05080767348408699, + 0.016855506226420403, + 0.285561740398407, + 0.07771508395671844, + 0.06848431378602982, + 0.27295663952827454, + 0.06255611777305603, + 0.03454965353012085, + 0.06874069571495056 + ], + "credit_map": [ + 0.013925384730100632, + 0.004765613470226526, + 0.00659734895452857, + 0.0036837186198681593, + 0.005540913436561823, + 0.006471441127359867, + 0.006646274589002132, + 0.014141885563731194, + 0.009489016607403755, + 0.004943149164319038, + 0.005389400292187929, + 0.0035111205652356148, + 0.005143498070538044, + 0.006896627601236105, + 0.0057548051699995995, + 0.009680054150521755, + 0.0019867713563144207, + 0.0013605597196146846, + 0.0015286720590665936, + 0.0010232499334961176, + 0.001347962417639792, + 0.0019342146115377545, + 0.003723766189068556, + 0.0039503080770373344, + 0.07004236429929733, + 0.01176745817065239, + 0.011182047426700592, + 0.012005343101918697, + 0.01018415205180645, + 0.007767327595502138, + 0.11516627669334412, + 0.04744676873087883, + 0.015191037207841873, + 0.004856193903833628, + 0.00552392890676856, + 0.004730899352580309, + 0.004178962204605341, + 0.004851543810218573, + 0.026633329689502716, + 0.01174918469041586, + 0.01325302291661501, + 0.0049522994086146355, + 0.006128713022917509, + 0.0045343004167079926, + 0.004463741555809975, + 0.0036493230145424604, + 0.021710535511374474, + 0.009792379103600979, + 0.04896245151758194, + 0.011971096508204937, + 0.008829755708575249, + 0.012598104774951935, + 0.010816596448421478, + 0.007681121584028006, + 0.11130394786596298, + 0.06079355254769325, + 0.00768074207007885, + 0.00843372568488121, + 0.004457515198737383, + 0.005448997486382723, + 0.010678974911570549, + 0.004782000556588173, + 0.012767544947564602, + 0.008306613191962242, + 0.004909875802695751, + 0.0019086120883002877, + 0.0034679535310715437, + 0.0019727323669940233, + 0.0020759175531566143, + 0.0038820572663098574, + 0.010074091143906116, + 0.006258413195610046, + 0.010343680158257484, + 0.004310394171625376, + 0.005532205104827881, + 0.003717269981279969, + 0.0035367263481020927, + 0.006938761565834284, + 0.02192007377743721, + 0.01244158111512661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.05493195354938507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slowly advancing medium shot, captured from a slight upward angle, we are transported deep into a misty Eastern wonderland. A gigantic leaf-tailed gecko rests peacefully on a vine entwined with an ancient tree, emitting a soft blue-green glow. Its skin, a complex camouflage pattern reminiscent of millennia-old bark, is covered in mottled dark green and earthy brown mosses, punctuated by patches of silvery-gray lichen, almost blending seamlessly into the background. Its enormous, luminous amber eyes, with vertical pupils, gleam with a glimmer of all-seeing wisdom. Its signature broad, leaf-like tail, shaped like a withered autumn leaf, with naturally broken edges and distinct veins, trembles ever so slightly, almost imperceptibly, in sync with the invisible aura flowing around it. A thick, milky fog slowly drifts through the forest. The vine's bioluminescence illuminates the surrounding mist and the shimmering particles of light floating in the air, casting shifting shadows on the gecko's body. The entire picture is presented in a cinematic 16:9 aspect ratio and 4K high-definition resolution, with every detail clearly visible, creating an atmosphere of extreme tranquility, solemnity and ancient mystery.", + "seed": 9133, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_11/index_8289_seed_9133_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01384862493723631, + "vlm_logprob": -0.01384862493723631, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.01384862493723631, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010287001729011536 + ], + "opposite_logprob_tokens": [ + -4.635286808013916 + ], + "logprob_reward": -0.010287001729011536, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.010287001729011536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24713511765003204, + 0.12257146090269089, + 0.04209514334797859, + 0.017992718145251274, + 0.040298108011484146, + 0.07145991921424866, + 0.033379536122083664, + 0.10417603701353073, + 0.2924712300300598, + 0.02842073142528534 + ], + "credit_map": [ + 0.016751296818256378, + 0.006078110076487064, + 0.0993901863694191, + 0.016840234398841858, + 0.010707591660320759, + 0.024600105360150337, + 0.008100531995296478, + 0.06466706842184067, + 0.011717444285750389, + 0.015545453876256943, + 0.018977917730808258, + 0.011721480637788773, + 0.014777339063584805, + 0.013444502837955952, + 0.02381112612783909, + 0.01257619820535183, + 0.008234148845076561, + 0.00422168243676424, + 0.006361092906445265, + 0.004850973375141621, + 0.003933924250304699, + 0.007608294952660799, + 0.0036608795635402203, + 0.0032241488806903362, + 0.0018769098678603768, + 0.0023872682359069586, + 0.0034361854195594788, + 0.002805129624903202, + 0.0014266427606344223, + 0.002110537374392152, + 0.002083827741444111, + 0.0018662188667804003, + 0.004465222824364901, + 0.002473436063155532, + 0.011373964138329029, + 0.003655759384855628, + 0.0028156982734799385, + 0.005790161434561014, + 0.004399567376822233, + 0.005324304569512606, + 0.008100101724267006, + 0.007572584319859743, + 0.011198455467820168, + 0.007688978686928749, + 0.006218275986611843, + 0.009996965527534485, + 0.010700052604079247, + 0.009984510019421577, + 0.0038182318676263094, + 0.0032964893616735935, + 0.0034318629186600447, + 0.0035279891453683376, + 0.003979903645813465, + 0.0043569342233240604, + 0.005945730954408646, + 0.005022395867854357, + 0.015313029289245605, + 0.010537357069551945, + 0.013106698170304298, + 0.010799054987728596, + 0.011360728181898594, + 0.010105988010764122, + 0.016177186742424965, + 0.01677599921822548, + 0.026147622615098953, + 0.023717721924185753, + 0.03165409341454506, + 0.030420714989304543, + 0.031033571809530258, + 0.029484709724783897, + 0.07071957737207413, + 0.04929322749376297, + 0.004199170507490635, + 0.004113932140171528, + 0.002955319359898567, + 0.002360180951654911, + 0.003397130174562335, + 0.004548210650682449, + 0.0028637098148465157, + 0.003983079921454191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.010287001729011536 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645840920507908 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008645840920507908, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645840920507908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2837460935115814, + 0.3174883723258972, + 0.059916697442531586, + 0.03909855708479881, + 0.024672165513038635, + 0.0455094538629055, + 0.030631037428975105, + 0.05140668898820877, + 0.09463731944561005, + 0.05289364606142044 + ], + "credit_map": [ + 0.020018411800265312, + 0.010396280325949192, + 0.10790197551250458, + 0.019142480567097664, + 0.014143208973109722, + 0.02669323794543743, + 0.011677865870296955, + 0.07377263903617859, + 0.0306064672768116, + 0.05274161696434021, + 0.04190812632441521, + 0.0299476720392704, + 0.040038228034973145, + 0.03687991946935654, + 0.053485285490751266, + 0.031881049275398254, + 0.011363335885107517, + 0.005395893007516861, + 0.014822477474808693, + 0.005986848380416632, + 0.004718602169305086, + 0.007431649602949619, + 0.004821965470910072, + 0.005375925917178392, + 0.003755986224859953, + 0.005545531865209341, + 0.008609771728515625, + 0.004746670834720135, + 0.0031548882834613323, + 0.005200181622058153, + 0.003740331856533885, + 0.004345194436609745, + 0.0030886975582689047, + 0.001850638771429658, + 0.0059433248825371265, + 0.0025904809590429068, + 0.0016839666059240699, + 0.0033506399486213923, + 0.0027609295211732388, + 0.0034034850541502237, + 0.005095191765576601, + 0.005323768127709627, + 0.006312906742095947, + 0.006111926399171352, + 0.0033116398844867945, + 0.00549729261547327, + 0.007899044081568718, + 0.005957684479653835, + 0.003338821465149522, + 0.003306744620203972, + 0.003288622945547104, + 0.004173910245299339, + 0.0034720830153673887, + 0.003283910918980837, + 0.005928113125264645, + 0.0038388303946703672, + 0.005785821937024593, + 0.006530983839184046, + 0.00526751484721899, + 0.005497914273291826, + 0.007253394927829504, + 0.005825485568493605, + 0.008162276819348335, + 0.007083295378834009, + 0.010093801654875278, + 0.01116874534636736, + 0.011172422207891941, + 0.009949933737516403, + 0.009920916520059109, + 0.011815761215984821, + 0.016491012647747993, + 0.01402471773326397, + 0.006000508088618517, + 0.01075773872435093, + 0.005356225185096264, + 0.004528403747826815, + 0.005712140817195177, + 0.005777623504400253, + 0.006020811386406422, + 0.008740193210542202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.008645840920507908 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782706387341022 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782706387341022, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782706387341022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41312867403030396, + 0.09868574887514114, + 0.06295064836740494, + 0.025056634098291397, + 0.025082748383283615, + 0.03224325180053711, + 0.033662792295217514, + 0.1268356442451477, + 0.14845481514930725, + 0.03389906883239746 + ], + "credit_map": [ + 0.029179014265537262, + 0.012797036208212376, + 0.16252024471759796, + 0.029559550806879997, + 0.018387116491794586, + 0.04028414189815521, + 0.014251569285988808, + 0.10614999383687973, + 0.008225463330745697, + 0.011078869923949242, + 0.018987221643328667, + 0.00932969618588686, + 0.01097799651324749, + 0.00952738057821989, + 0.01923542469739914, + 0.011323685757815838, + 0.010203549638390541, + 0.006841519847512245, + 0.011775042861700058, + 0.005911251995712519, + 0.005709535907953978, + 0.011787976138293743, + 0.005212404299527407, + 0.005509358365088701, + 0.0026468217838555574, + 0.0031227977015078068, + 0.004593370482325554, + 0.0029597152024507523, + 0.0023925418499857187, + 0.003598107025027275, + 0.0026185978204011917, + 0.0031246794387698174, + 0.0025282076094299555, + 0.0029979385435581207, + 0.005197898950427771, + 0.0025807812344282866, + 0.0026005874387919903, + 0.003220781683921814, + 0.0030429845210164785, + 0.0029135674703866243, + 0.0036338584031909704, + 0.0034618612844496965, + 0.00381206045858562, + 0.0033064705785363913, + 0.0036940407007932663, + 0.003386792028322816, + 0.006614734418690205, + 0.004333433695137501, + 0.005338010378181934, + 0.0037816506810486317, + 0.004198904614895582, + 0.0032351575791835785, + 0.0037044985219836235, + 0.0040009524673223495, + 0.0044158571399748325, + 0.004987761378288269, + 0.01872512884438038, + 0.013686499558389187, + 0.013252675533294678, + 0.01285080797970295, + 0.014930346049368382, + 0.01225712988525629, + 0.022202758118510246, + 0.018930289894342422, + 0.01795158162713051, + 0.015707235783338547, + 0.021524880081415176, + 0.017291150987148285, + 0.011692513711750507, + 0.018477993085980415, + 0.027613602578639984, + 0.018195865675807, + 0.003130862955003977, + 0.005267800763249397, + 0.0037300926633179188, + 0.0030872272327542305, + 0.006332214456051588, + 0.0035865535028278828, + 0.0038058224599808455, + 0.004958493169397116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.009782706387341022 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634073793888092 + ], + "opposite_logprob_tokens": [ + -3.6513407230377197 + ], + "logprob_reward": -0.02634073793888092, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634073793888092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32942503690719604, + 0.1210852712392807, + 0.04526009038090706, + 0.039187848567962646, + 0.033276528120040894, + 0.045158594846725464, + 0.07857382297515869, + 0.1244208887219429, + 0.13414183259010315, + 0.04947013407945633 + ], + "credit_map": [ + 0.0222318135201931, + 0.014654154889285564, + 0.13036613166332245, + 0.02078498899936676, + 0.016321849077939987, + 0.030495138838887215, + 0.009835404343903065, + 0.08473554998636246, + 0.017909513786435127, + 0.016700081527233124, + 0.01984255202114582, + 0.011412860825657845, + 0.01589050516486168, + 0.011840690858662128, + 0.015906505286693573, + 0.011582549661397934, + 0.0057804216630756855, + 0.0067638070322573185, + 0.0053497739136219025, + 0.005805536638945341, + 0.005068831145763397, + 0.007065923884510994, + 0.004974690731614828, + 0.004451100714504719, + 0.003853621892631054, + 0.0063692061230540276, + 0.006966985296458006, + 0.00434856116771698, + 0.004549866076558828, + 0.004204329568892717, + 0.004798586945980787, + 0.004096689634025097, + 0.0038019914645701647, + 0.003003096440806985, + 0.005372611340135336, + 0.0033445053268224, + 0.003257686970755458, + 0.0057197571732103825, + 0.003565762657672167, + 0.005211115349084139, + 0.003939284011721611, + 0.004450142849236727, + 0.004781599156558514, + 0.0058263149112463, + 0.004604383837431669, + 0.0074835289269685745, + 0.007939085364341736, + 0.006134251598268747, + 0.009591583162546158, + 0.007310971152037382, + 0.00873379223048687, + 0.009863260202109814, + 0.009930024854838848, + 0.010084244422614574, + 0.012116236612200737, + 0.0109436996281147, + 0.014390540309250355, + 0.013654128648340702, + 0.010867487639188766, + 0.014088321477174759, + 0.01861877366900444, + 0.008441238664090633, + 0.025686919689178467, + 0.01867346093058586, + 0.011560939252376556, + 0.013731015846133232, + 0.012655047699809074, + 0.011322579346597195, + 0.019057728350162506, + 0.013189922086894512, + 0.027116425335407257, + 0.02550816349685192, + 0.0055290586315095425, + 0.006997876334935427, + 0.004498699214309454, + 0.004708576947450638, + 0.006051270756870508, + 0.00726965069770813, + 0.006805983372032642, + 0.007609013468027115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.02634073793888092 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186837710440159 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.014186837710440159, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186837710440159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031832315027713776, + 0.43253493309020996, + 0.07196585088968277, + 0.07542172074317932, + 0.03254516050219536, + 0.01096119824796915, + 0.029020968824625015, + 0.13220083713531494, + 0.15489286184310913, + 0.028624219819903374 + ], + "credit_map": [ + 0.004302504472434521, + 0.003922174219042063, + 0.005434674676507711, + 0.0037368983030319214, + 0.0036530904471874237, + 0.003428034484386444, + 0.003467430593445897, + 0.003887507598847151, + 0.045129552483558655, + 0.07651497423648834, + 0.060023605823516846, + 0.0393841415643692, + 0.05837474763393402, + 0.05026359483599663, + 0.062118228524923325, + 0.040726061910390854, + 0.011147845536470413, + 0.005511609837412834, + 0.02043750509619713, + 0.007172625977545977, + 0.005353409796953201, + 0.008936723694205284, + 0.006662299390882254, + 0.006743817590177059, + 0.008881887421011925, + 0.005080935079604387, + 0.015576886013150215, + 0.007686900440603495, + 0.00547025864943862, + 0.017134500667452812, + 0.0073543754406273365, + 0.008235970512032509, + 0.0036371578462421894, + 0.002817206084728241, + 0.006074842531234026, + 0.002969169057905674, + 0.003097090870141983, + 0.006118131335824728, + 0.0037773943040519953, + 0.0040541645139455795, + 0.0012397743994370103, + 0.0011886820429936051, + 0.001310754450969398, + 0.0015413357177749276, + 0.0011294191936030984, + 0.001121298409998417, + 0.0020400243811309338, + 0.0013899080222472548, + 0.003312724409624934, + 0.002400477882474661, + 0.003109759883955121, + 0.0037339823320508003, + 0.003547860309481621, + 0.00366872176527977, + 0.005007988307625055, + 0.004239450674504042, + 0.015120627358555794, + 0.014646273106336594, + 0.012887847609817982, + 0.014475584961473942, + 0.018099311739206314, + 0.010289615020155907, + 0.027555422857403755, + 0.019126132130622864, + 0.017636490985751152, + 0.01662059873342514, + 0.018800197169184685, + 0.015163185074925423, + 0.015325000509619713, + 0.019190844148397446, + 0.02923785150051117, + 0.022918671369552612, + 0.0042520323768258095, + 0.0030765943229198456, + 0.002949570072814822, + 0.0030875930096954107, + 0.003741469467058778, + 0.004904852248728275, + 0.0030862956773489714, + 0.0035258133430033922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.014186837710440159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slowly advancing medium shot, captured from a slight upward angle, we are transported deep into a misty Eastern wonderland. A gigantic leaf-tailed gecko rests peacefully on a vine entwined with an ancient tree, emitting a soft blue-green glow. Its skin, a complex camouflage pattern reminiscent of millennia-old bark, is covered in mottled dark green and earthy brown mosses, punctuated by patches of silvery-gray lichen, almost blending seamlessly into the background. Its enormous, luminous amber eyes, with vertical pupils, gleam with a glimmer of all-seeing wisdom. Its signature broad, leaf-like tail, shaped like a withered autumn leaf, with naturally broken edges and distinct veins, trembles ever so slightly, almost imperceptibly, in sync with the invisible aura flowing around it. A thick, milky fog slowly drifts through the forest. The vine's bioluminescence illuminates the surrounding mist and the shimmering particles of light floating in the air, casting shifting shadows on the gecko's body. The entire picture is presented in a cinematic 16:9 aspect ratio and 4K high-definition resolution, with every detail clearly visible, creating an atmosphere of extreme tranquility, solemnity and ancient mystery.", + "seed": 9134, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_11/index_8289_seed_9134_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02302344683557749, + "vlm_logprob": -0.02302344683557749, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.02302344683557749, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019379209727048874 + ], + "opposite_logprob_tokens": [ + -4.019379138946533 + ], + "logprob_reward": -0.019379209727048874, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019379209727048874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1958404928445816, + 0.11351358890533447, + 0.081269770860672, + 0.06754034757614136, + 0.0746702179312706, + 0.05423735827207565, + 0.16899345815181732, + 0.12496485561132431, + 0.0639638751745224, + 0.05500604212284088 + ], + "credit_map": [ + 0.02160545252263546, + 0.02314499206840992, + 0.014982485212385654, + 0.018417276442050934, + 0.027696765959262848, + 0.034067314118146896, + 0.02580048330128193, + 0.030125726014375687, + 0.011357780545949936, + 0.008366906084120274, + 0.009704232215881348, + 0.011046410538256168, + 0.010666615329682827, + 0.019598061218857765, + 0.020404763519763947, + 0.022368809208273888, + 0.013410180807113647, + 0.008742288686335087, + 0.00968500692397356, + 0.012179219163954258, + 0.007829762063920498, + 0.011249477975070477, + 0.008745105005800724, + 0.009428734891116619, + 0.013831929303705692, + 0.011135740205645561, + 0.0073469714261591434, + 0.008351759985089302, + 0.007122781593352556, + 0.006329055409878492, + 0.004618850536644459, + 0.008803257718682289, + 0.008391929790377617, + 0.007906167767941952, + 0.00844944454729557, + 0.013239556923508644, + 0.005811972077935934, + 0.007530640810728073, + 0.008231692016124725, + 0.015108814463019371, + 0.004117440897971392, + 0.007085994817316532, + 0.006699632853269577, + 0.0053699202835559845, + 0.008309460245072842, + 0.007501693442463875, + 0.007345093879848719, + 0.007808121386915445, + 0.015833009034395218, + 0.02971404604613781, + 0.011664316989481449, + 0.01759854517877102, + 0.014172482304275036, + 0.03924485296010971, + 0.013965887948870659, + 0.026800312101840973, + 0.012802879326045513, + 0.012095070444047451, + 0.012248109094798565, + 0.013386983424425125, + 0.014489728026092052, + 0.018088635057210922, + 0.023650823161005974, + 0.018202630802989006, + 0.006504193879663944, + 0.005632459651678801, + 0.004997375886887312, + 0.006912780459970236, + 0.005967258010059595, + 0.016669977456331253, + 0.006734304595738649, + 0.010545521043241024, + 0.006683249492198229, + 0.0066930693574249744, + 0.005044800695031881, + 0.00539183197543025, + 0.00576558755710721, + 0.007621117401868105, + 0.009652425535023212, + 0.008153961040079594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.019379209727048874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330053597688675 + ], + "opposite_logprob_tokens": [ + -3.7733006477355957 + ], + "logprob_reward": -0.02330053597688675, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330053597688675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11323948204517365, + 0.2569092810153961, + 0.05426013469696045, + 0.14631853997707367, + 0.05882732942700386, + 0.08155516535043716, + 0.03839988633990288, + 0.15824687480926514, + 0.03913385048508644, + 0.05310942232608795 + ], + "credit_map": [ + 0.00904452707618475, + 0.012490306980907917, + 0.008204018697142601, + 0.011968903243541718, + 0.021304678171873093, + 0.021367603912949562, + 0.015512249432504177, + 0.013347199186682701, + 0.02925451472401619, + 0.025049734860658646, + 0.03434177488088608, + 0.04926745221018791, + 0.0374937541782856, + 0.02887357957661152, + 0.02684422768652439, + 0.02578422613441944, + 0.005129902623593807, + 0.004851819947361946, + 0.008097768761217594, + 0.011906764470040798, + 0.006234303582459688, + 0.0052088783122599125, + 0.006201834883540869, + 0.006628865841776133, + 0.017647970467805862, + 0.021001150831580162, + 0.01626758463680744, + 0.01757482811808586, + 0.016789739951491356, + 0.02300136908888817, + 0.01600591652095318, + 0.01802997477352619, + 0.005846552085131407, + 0.004532466642558575, + 0.007703160867094994, + 0.00888078473508358, + 0.004032410681247711, + 0.009207296185195446, + 0.005783171858638525, + 0.012841488234698772, + 0.009381171315908432, + 0.015204677358269691, + 0.006437376607209444, + 0.007177794352173805, + 0.009487300179898739, + 0.011978057213127613, + 0.010533645749092102, + 0.011355137452483177, + 0.0037115246523171663, + 0.008235466666519642, + 0.0036715471651405096, + 0.005619765724986792, + 0.005557315424084663, + 0.004363455343991518, + 0.003079075366258621, + 0.004161735996603966, + 0.014446313492953777, + 0.01527690701186657, + 0.01605784334242344, + 0.012782728299498558, + 0.026309680193662643, + 0.023706376552581787, + 0.02741318754851818, + 0.02225383371114731, + 0.004062710329890251, + 0.0046372259967029095, + 0.0035700907465070486, + 0.004044191911816597, + 0.004493779502809048, + 0.008029373362660408, + 0.0052576614543795586, + 0.005038819741457701, + 0.005830294452607632, + 0.004833684302866459, + 0.004104720428586006, + 0.005529386457055807, + 0.003974478226155043, + 0.012555396184325218, + 0.00806296057999134, + 0.008218502625823021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.02330053597688675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110772168263793 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.0110772168263793, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110772168263793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1510743349790573, + 0.1676318198442459, + 0.12006475031375885, + 0.0900036171078682, + 0.10426752269268036, + 0.06601335108280182, + 0.09598744660615921, + 0.10204591602087021, + 0.05746963620185852, + 0.04544166848063469 + ], + "credit_map": [ + 0.012256277725100517, + 0.012291188351809978, + 0.013352607376873493, + 0.017045369371771812, + 0.03129122406244278, + 0.02619888447225094, + 0.02169007621705532, + 0.016948696225881577, + 0.027694741263985634, + 0.019981475546956062, + 0.026674117892980576, + 0.03480222448706627, + 0.01383446529507637, + 0.011033623479306698, + 0.019681699573993683, + 0.013929454609751701, + 0.01153349969536066, + 0.01097178179770708, + 0.01849958673119545, + 0.022589020431041718, + 0.012501001358032227, + 0.017398521304130554, + 0.014628736302256584, + 0.011942592449486256, + 0.01310714241117239, + 0.010296713560819626, + 0.011745575815439224, + 0.013836869038641453, + 0.011271020397543907, + 0.0092548793181777, + 0.007628887891769409, + 0.012862521223723888, + 0.010110646486282349, + 0.017167584970593452, + 0.01113855466246605, + 0.014728136360645294, + 0.008883916772902012, + 0.01449414249509573, + 0.010148150846362114, + 0.017596378922462463, + 0.008623712696135044, + 0.007626305799931288, + 0.006314218044281006, + 0.006784943398088217, + 0.006918524857610464, + 0.012069011107087135, + 0.00887388363480568, + 0.00880274549126625, + 0.0057488069869577885, + 0.008521219715476036, + 0.010615186765789986, + 0.00885243620723486, + 0.011928200721740723, + 0.026235416531562805, + 0.008998468518257141, + 0.015087708830833435, + 0.007815777324140072, + 0.008794188499450684, + 0.007912029512226582, + 0.010918129235506058, + 0.011038596741855145, + 0.020265089347958565, + 0.018111664801836014, + 0.01719043217599392, + 0.004867196548730135, + 0.005875640083104372, + 0.0061857448890805244, + 0.005891673732548952, + 0.007726599927991629, + 0.013069532811641693, + 0.005256849806755781, + 0.00859639048576355, + 0.006087467540055513, + 0.004640108905732632, + 0.004561110399663448, + 0.004830699879676104, + 0.005024591460824013, + 0.006965680979192257, + 0.005789196118712425, + 0.007542806211858988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0110772168263793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380859412252903 + ], + "opposite_logprob_tokens": [ + -3.2880859375 + ], + "logprob_reward": -0.0380859412252903, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380859412252903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04253920540213585, + 0.09212608635425568, + 0.0684160590171814, + 0.1679007112979889, + 0.045745279639959335, + 0.06710080802440643, + 0.09590578079223633, + 0.31335827708244324, + 0.0625142753124237, + 0.044393472373485565 + ], + "credit_map": [ + 0.004832522943615913, + 0.004433739930391312, + 0.005370378494262695, + 0.005053170956671238, + 0.007306648418307304, + 0.0055727772414684296, + 0.005410289391875267, + 0.004559677094221115, + 0.015254261903464794, + 0.009893521666526794, + 0.013880239799618721, + 0.01591361127793789, + 0.009921611286699772, + 0.009486991912126541, + 0.010088343173265457, + 0.00768751185387373, + 0.0077574001625180244, + 0.006432505324482918, + 0.01164126768708229, + 0.02009599655866623, + 0.004378636367619038, + 0.005162087269127369, + 0.006907985545694828, + 0.006040181964635849, + 0.0270779337733984, + 0.022901274263858795, + 0.01937897875905037, + 0.020785177126526833, + 0.023944346234202385, + 0.017188623547554016, + 0.013652266934514046, + 0.022972097620368004, + 0.005028276704251766, + 0.005449289921671152, + 0.005012162029743195, + 0.00647661741822958, + 0.005380154587328434, + 0.006020253524184227, + 0.004133324138820171, + 0.008245200850069523, + 0.007588500622659922, + 0.011568691581487656, + 0.007127540186047554, + 0.008101742714643478, + 0.007240680977702141, + 0.008547651581466198, + 0.006130964029580355, + 0.010795039124786854, + 0.007690239232033491, + 0.014152182266116142, + 0.008117519319057465, + 0.008861244656145573, + 0.008726528845727444, + 0.02497778832912445, + 0.008087060414254665, + 0.015293224714696407, + 0.02794652059674263, + 0.02685847133398056, + 0.029909692704677582, + 0.026023976504802704, + 0.03961845487356186, + 0.0390893928706646, + 0.07485835999250412, + 0.04905340448021889, + 0.007557840086519718, + 0.006749278865754604, + 0.006189315114170313, + 0.007140986621379852, + 0.006961315404623747, + 0.01337802316993475, + 0.005416556727141142, + 0.009120956994593143, + 0.0065413848496973515, + 0.004496335983276367, + 0.0033466957975178957, + 0.00483194412663579, + 0.0048286523669958115, + 0.007842429913580418, + 0.005453153047710657, + 0.007052876986563206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.0380859412252903 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327433042228222 + ], + "opposite_logprob_tokens": [ + -3.7732744216918945 + ], + "logprob_reward": -0.02327433042228222, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327433042228222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0963907316327095, + 0.06780792027711868, + 0.0782703310251236, + 0.15864986181259155, + 0.1583137959241867, + 0.13944612443447113, + 0.1468900740146637, + 0.04976579174399376, + 0.029397230595350266, + 0.07506812363862991 + ], + "credit_map": [ + 0.009276444092392921, + 0.009281734935939312, + 0.010756716132164001, + 0.015533502213656902, + 0.013173023238778114, + 0.01245053019374609, + 0.012687264941632748, + 0.013231517747044563, + 0.011978371068835258, + 0.007304037921130657, + 0.00789622962474823, + 0.009517536498606205, + 0.008496706373989582, + 0.006044091656804085, + 0.009133449755609035, + 0.007437499240040779, + 0.009547398425638676, + 0.005827534478157759, + 0.00981476716697216, + 0.007360348477959633, + 0.008056501857936382, + 0.013493041507899761, + 0.012325935997068882, + 0.011844800785183907, + 0.0239497609436512, + 0.023446455597877502, + 0.017771653831005096, + 0.02112356759607792, + 0.021019544452428818, + 0.018754743039608, + 0.01345604844391346, + 0.01912808045744896, + 0.014865677803754807, + 0.012202258221805096, + 0.01751825399696827, + 0.022433212026953697, + 0.012548407539725304, + 0.03654663264751434, + 0.01712050475180149, + 0.025078855454921722, + 0.013212216086685658, + 0.013304672203958035, + 0.012382269836962223, + 0.011717554181814194, + 0.013981446623802185, + 0.033964045345783234, + 0.017238665372133255, + 0.023645253852009773, + 0.009625881910324097, + 0.018135540187358856, + 0.012769094668328762, + 0.014013751409947872, + 0.01579410955309868, + 0.034174323081970215, + 0.018602170050144196, + 0.02377520501613617, + 0.004115819465368986, + 0.004601387307047844, + 0.004689098335802555, + 0.004536351654678583, + 0.007075399160385132, + 0.006314049940556288, + 0.01113542914390564, + 0.007298254407942295, + 0.0027596442960202694, + 0.0025575377512723207, + 0.0029073385521769524, + 0.003866502782329917, + 0.00267417193390429, + 0.007302524987608194, + 0.00303089153021574, + 0.004298618528991938, + 0.008698003366589546, + 0.006647081580013037, + 0.004924907814711332, + 0.006183227989822626, + 0.005931893363595009, + 0.02021438255906105, + 0.010786882601678371, + 0.011681740172207355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.02327433042228222 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a slowly advancing medium shot, captured from a slight upward angle, we are transported deep into a misty Eastern wonderland. A gigantic leaf-tailed gecko rests peacefully on a vine entwined with an ancient tree, emitting a soft blue-green glow. Its skin, a complex camouflage pattern reminiscent of millennia-old bark, is covered in mottled dark green and earthy brown mosses, punctuated by patches of silvery-gray lichen, almost blending seamlessly into the background. Its enormous, luminous amber eyes, with vertical pupils, gleam with a glimmer of all-seeing wisdom. Its signature broad, leaf-like tail, shaped like a withered autumn leaf, with naturally broken edges and distinct veins, trembles ever so slightly, almost imperceptibly, in sync with the invisible aura flowing around it. A thick, milky fog slowly drifts through the forest. The vine's bioluminescence illuminates the surrounding mist and the shimmering particles of light floating in the air, casting shifting shadows on the gecko's body. The entire picture is presented in a cinematic 16:9 aspect ratio and 4K high-definition resolution, with every detail clearly visible, creating an atmosphere of extreme tranquility, solemnity and ancient mystery.", + "seed": 9135, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_11/index_8289_seed_9135_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.030165884643793106, + "vlm_logprob": -0.030165884643793106, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.030165884643793106, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.056135546416044235 + ], + "opposite_logprob_tokens": [ + -2.931135654449463 + ], + "logprob_reward": -0.056135546416044235, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.056135546416044235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12657053768634796, + 0.06622987240552902, + 0.14214631915092468, + 0.2713654041290283, + 0.12587137520313263, + 0.06980345398187637, + 0.029225260019302368, + 0.0839393362402916, + 0.030624736100435257, + 0.05422366037964821 + ], + "credit_map": [ + 0.013668768107891083, + 0.014197379350662231, + 0.02627827599644661, + 0.02280864678323269, + 0.009789717383682728, + 0.010380661115050316, + 0.01631852611899376, + 0.01312857586890459, + 0.011290735565125942, + 0.011660524643957615, + 0.008294453844428062, + 0.009561438113451004, + 0.005400273948907852, + 0.007118606474250555, + 0.007291450165212154, + 0.00561239430680871, + 0.03107965551316738, + 0.01130773313343525, + 0.0469801127910614, + 0.012215344235301018, + 0.014045581221580505, + 0.0069890557788312435, + 0.01039994228631258, + 0.009128917008638382, + 0.037127554416656494, + 0.0132330022752285, + 0.11134445667266846, + 0.01467834785580635, + 0.023969225585460663, + 0.012293036095798016, + 0.014174330048263073, + 0.044545482844114304, + 0.024653960019350052, + 0.007253090385347605, + 0.041971806436777115, + 0.01425796840339899, + 0.015145408920943737, + 0.005919793155044317, + 0.009513567201793194, + 0.007155795581638813, + 0.012647230178117752, + 0.00919459480792284, + 0.007448872085660696, + 0.01033737976104021, + 0.007880325429141521, + 0.005537910386919975, + 0.008826366625726223, + 0.007930787280201912, + 0.0035176677629351616, + 0.0041766357608139515, + 0.00296856090426445, + 0.0030777284409850836, + 0.003906444180756807, + 0.0033044267911463976, + 0.00432467320933938, + 0.003949126694351435, + 0.02751856856048107, + 0.005723483860492706, + 0.005392325110733509, + 0.002986965235322714, + 0.026028454303741455, + 0.006904625799506903, + 0.005267336964607239, + 0.004117585252970457, + 0.004691264126449823, + 0.003512633964419365, + 0.0019395517883822322, + 0.002002021996304393, + 0.009431712329387665, + 0.002064760774374008, + 0.002509069861844182, + 0.004473724868148565, + 0.006332503631711006, + 0.009961327537894249, + 0.006361710373312235, + 0.00585639663040638, + 0.010311590507626534, + 0.00517737353220582, + 0.004569216165691614, + 0.005653546191751957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.056135546416044235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078222543001175 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.016078222543001175, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078222543001175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1809512823820114, + 0.051745008677244186, + 0.16548407077789307, + 0.2009497433900833, + 0.14006417989730835, + 0.08801867812871933, + 0.027361761778593063, + 0.07134601473808289, + 0.021115044131875038, + 0.052964191883802414 + ], + "credit_map": [ + 0.02182324230670929, + 0.023717360571026802, + 0.028788868337869644, + 0.038604315370321274, + 0.017423778772354126, + 0.011815805919468403, + 0.020552320405840874, + 0.018225600942969322, + 0.007841612212359905, + 0.0066352784633636475, + 0.009671962819993496, + 0.0067940810695290565, + 0.004521639551967382, + 0.005119789391756058, + 0.005974819418042898, + 0.005185830872505903, + 0.028071697801351547, + 0.015955792739987373, + 0.040952589362859726, + 0.023489510640501976, + 0.019656503573060036, + 0.010166900232434273, + 0.01578645408153534, + 0.011404612101614475, + 0.027349920943379402, + 0.011844122782349586, + 0.07650620490312576, + 0.010494676418602467, + 0.02531157061457634, + 0.008420395664870739, + 0.010172867216169834, + 0.0308499988168478, + 0.03780386224389076, + 0.0069634513929486275, + 0.056993044912815094, + 0.010479006916284561, + 0.007948344573378563, + 0.004070911090821028, + 0.008114580996334553, + 0.007690977770835161, + 0.01521818432956934, + 0.011791804805397987, + 0.009340745396912098, + 0.013246111571788788, + 0.011066622100770473, + 0.007643437944352627, + 0.00969532411545515, + 0.010016451589763165, + 0.0033681094646453857, + 0.0030041811987757683, + 0.00317956879734993, + 0.003769900184124708, + 0.0027262798976153135, + 0.0034407596103847027, + 0.00446629012003541, + 0.0034066748339682817, + 0.01959085278213024, + 0.006545496638864279, + 0.004775103647261858, + 0.00381943816319108, + 0.01958523504436016, + 0.006667529232800007, + 0.004957340192049742, + 0.005405028350651264, + 0.0023713719565421343, + 0.0033359096851199865, + 0.0025073273573070765, + 0.003189896233379841, + 0.003167487448081374, + 0.0019121652003377676, + 0.002688497304916382, + 0.001942390576004982, + 0.0053996010683476925, + 0.006227725185453892, + 0.007629651576280594, + 0.0077163223177194595, + 0.007777730002999306, + 0.006258197594434023, + 0.005809512455016375, + 0.006145453546196222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.016078222543001175 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194712042808533 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194712042808533, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194712042808533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13837243616580963, + 0.05962619557976723, + 0.09751086682081223, + 0.20298530161380768, + 0.21151810884475708, + 0.024785256013274193, + 0.027064792811870575, + 0.1678590327501297, + 0.029672708362340927, + 0.040605328977108 + ], + "credit_map": [ + 0.021687842905521393, + 0.016559738665819168, + 0.03208896517753601, + 0.013201167806982994, + 0.02334803342819214, + 0.01047853659838438, + 0.014149649068713188, + 0.00685849180445075, + 0.009939099662005901, + 0.010379682295024395, + 0.006087204907089472, + 0.009772386401891708, + 0.006903128698468208, + 0.005942419171333313, + 0.005947273224592209, + 0.004654998425394297, + 0.020595645532011986, + 0.007454739883542061, + 0.0326160192489624, + 0.00986268650740385, + 0.009342384524643421, + 0.004788204096257687, + 0.006867570336908102, + 0.0059836129657924175, + 0.032579388469457626, + 0.008323176763951778, + 0.08268669247627258, + 0.012676275335252285, + 0.018062978982925415, + 0.008820870891213417, + 0.010239504277706146, + 0.029596393927931786, + 0.05097067356109619, + 0.010746902786195278, + 0.07926677167415619, + 0.022012989968061447, + 0.01541099976748228, + 0.006793146021664143, + 0.01427800115197897, + 0.012038640677928925, + 0.004076812416315079, + 0.003948891069740057, + 0.0025431481190025806, + 0.0028947158716619015, + 0.003080586437135935, + 0.00209787511266768, + 0.003142752917483449, + 0.0030004738364368677, + 0.0047855922020971775, + 0.003647282486781478, + 0.002040879102423787, + 0.004528365563601255, + 0.0032912406604737043, + 0.0025741495192050934, + 0.004497837740927935, + 0.0016994447214528918, + 0.04071786627173424, + 0.031183592975139618, + 0.007738076150417328, + 0.005269092041999102, + 0.05925978720188141, + 0.011066590435802937, + 0.008603759109973907, + 0.004020263906568289, + 0.0038223247975111008, + 0.0036088312044739723, + 0.0025734053924679756, + 0.0039095086976885796, + 0.00764217134565115, + 0.0022934183944016695, + 0.003235402749851346, + 0.00258764554746449, + 0.004598068073391914, + 0.007122273091226816, + 0.0040474203415215015, + 0.005136770196259022, + 0.008639891631901264, + 0.003791996045038104, + 0.0037754245568066835, + 0.0034934855066239834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.014194712042808533 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808984160423279 + ], + "opposite_logprob_tokens": [ + -3.2880897521972656 + ], + "logprob_reward": -0.03808984160423279, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808984160423279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10310108959674835, + 0.09291494637727737, + 0.0649525448679924, + 0.2619785666465759, + 0.17925073206424713, + 0.0317828468978405, + 0.03241243585944176, + 0.13199417293071747, + 0.053630780428647995, + 0.04798182472586632 + ], + "credit_map": [ + 0.011906497180461884, + 0.010010031051933765, + 0.0379658006131649, + 0.020103394985198975, + 0.006113582290709019, + 0.004020861815661192, + 0.008367096073925495, + 0.004613830707967281, + 0.012014343403279781, + 0.012060360983014107, + 0.013585259206593037, + 0.013822680339217186, + 0.014123007655143738, + 0.010081804357469082, + 0.01021831575781107, + 0.007009175606071949, + 0.013051203452050686, + 0.006382617633789778, + 0.008106743916869164, + 0.008370479568839073, + 0.00800349935889244, + 0.00711526395753026, + 0.00773661769926548, + 0.006186123006045818, + 0.03412916138768196, + 0.019412294030189514, + 0.09608041495084763, + 0.022433141246438026, + 0.031090276315808296, + 0.010451915673911572, + 0.01443310733884573, + 0.03394826129078865, + 0.049267880618572235, + 0.008176032453775406, + 0.06835534423589706, + 0.01645817421376705, + 0.009984846226871014, + 0.004915313329547644, + 0.011377374641597271, + 0.010715759359300137, + 0.004773390479385853, + 0.0059142629615962505, + 0.0037667942233383656, + 0.004609519615769386, + 0.0049139452166855335, + 0.0022567291744053364, + 0.003897745395079255, + 0.0016504593659192324, + 0.003687022253870964, + 0.00461017107591033, + 0.003334715263918042, + 0.004881064407527447, + 0.0029935124330222607, + 0.0034476916771382093, + 0.0052065192721784115, + 0.004251738078892231, + 0.009719247929751873, + 0.05622656270861626, + 0.004580690059810877, + 0.005875800736248493, + 0.03639807552099228, + 0.006179737392812967, + 0.009979956783354282, + 0.003034101566299796, + 0.005499725230038166, + 0.0075731463730335236, + 0.006450682878494263, + 0.005989523604512215, + 0.007789143361151218, + 0.0071745142340660095, + 0.005924223456531763, + 0.007229818031191826, + 0.004697086755186319, + 0.005504943430423737, + 0.006709768902510405, + 0.0059372722171247005, + 0.006821152754127979, + 0.00610912824049592, + 0.005910167936235666, + 0.006292302627116442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.03808984160423279 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a gigantic leaf-tailed gecko visible resting on a vine entwined with an ancient tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the gecko emit a soft blue-green glow and have luminous amber eyes with vertical pupils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the gecko's skin covered in a camouflage pattern of mottled dark green, earthy brown mosses, and silvery-gray lichen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the gecko's broad, leaf-like tail with broken edges and distinct veins visible trembling slightly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slowly advancing medium shot captured from a slight upward angle with a thick milky fog drifting through the forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263311006128788 + ], + "opposite_logprob_tokens": [ + -3.6513311862945557 + ], + "logprob_reward": -0.0263311006128788, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263311006128788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09279008954763412, + 0.055024418979883194, + 0.12537598609924316, + 0.16160620748996735, + 0.0979938805103302, + 0.04685598984360695, + 0.04833704233169556, + 0.31639960408210754, + 0.020069606602191925, + 0.035547222942113876 + ], + "credit_map": [ + 0.014171790331602097, + 0.012652360834181309, + 0.012374870479106903, + 0.01730896532535553, + 0.009649953804910183, + 0.007753346581012011, + 0.011579159647226334, + 0.007299633696675301, + 0.006318349391222, + 0.007450134493410587, + 0.009403514675796032, + 0.007680597715079784, + 0.006262348964810371, + 0.004783515352755785, + 0.005442087538540363, + 0.007683870382606983, + 0.021684445440769196, + 0.01444348506629467, + 0.012197837233543396, + 0.015228386037051678, + 0.021314354613423347, + 0.011094856075942516, + 0.013709654100239277, + 0.015702959150075912, + 0.024734793230891228, + 0.008424941450357437, + 0.06396222114562988, + 0.009247368201613426, + 0.017194172367453575, + 0.007399471942335367, + 0.0089630838483572, + 0.021680139005184174, + 0.023313745856285095, + 0.004428595770150423, + 0.03898772597312927, + 0.009458123706281185, + 0.006723783444613218, + 0.003051938721910119, + 0.006061939522624016, + 0.005968013778328896, + 0.00993066281080246, + 0.006017627194523811, + 0.00408547930419445, + 0.007055999711155891, + 0.006071038078516722, + 0.003749354276806116, + 0.00518875801935792, + 0.004757068585604429, + 0.009056073613464832, + 0.013923143036663532, + 0.0035106572322547436, + 0.0036042602732777596, + 0.006071099080145359, + 0.004407287575304508, + 0.00421628775075078, + 0.003548226784914732, + 0.08203733712434769, + 0.04665808379650116, + 0.014457275159657001, + 0.008971969597041607, + 0.11450066417455673, + 0.02383565530180931, + 0.016019348055124283, + 0.009919234551489353, + 0.0024489148054271936, + 0.0016502399230375886, + 0.0025765455793589354, + 0.002594027202576399, + 0.002300745341926813, + 0.0024658639449626207, + 0.00276168598793447, + 0.00327158160507679, + 0.004101857077330351, + 0.006606812123209238, + 0.003761340631172061, + 0.004558261018246412, + 0.007233033888041973, + 0.003517041215673089, + 0.002905889181420207, + 0.0028629873413592577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.0263311006128788 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_7/index_tensor([8938, 8938])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_7/index_tensor([8938, 8938])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..64aa75ad466d1daf6d01d9dabc382b70e133e8b6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_7/index_tensor([8938, 8938])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping wide-angle shot slowly sweeps across a rain-swept, rain-soaked urban canyon, capturing scenes of catastrophic destruction within a dark, dystopian sci-fi aesthetic. From a turbulent, inky-black sky, massive, crystalline hailstones, some the size of smaller boulders with jagged, faceted surfaces, descend at terrifying speed. They impact with explosive force, violently shattering the ornate, neo-futuristic facades of towering skyscrapers. The camera follows one such impact as a hailstone smashes into a gargoyle-adorned windowsill, sending a shower of black stone and gleaming ice shards into the cracks below. At street level, the sleek, aerodynamic shells of advanced vehicles are systematically crushed and dented by the relentless ice assault. Large, unstable holographic advertisements flicker erratically between buildings, their vibrant cyan and purplish-red light casting chaotic patterns, shifting reflections across the wet asphalt, translucent hailstones, and the smoke rising from short-circuited machinery, creating an atmosphere of overwhelming chaos both terrifying and captivating.", + "seed": 15551, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_7/index_8938_seed_15551_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04131946330890059, + "vlm_logprob": -0.04131946330890059, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.04131946330890059, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011299780569970608 + ], + "opposite_logprob_tokens": [ + -4.6362996101379395 + ], + "logprob_reward": -0.011299780569970608, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011299780569970608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25226062536239624, + 0.08012876659631729, + 0.1386575549840927, + 0.0870782881975174, + 0.03646829351782799, + 0.04335976764559746, + 0.0743403211236, + 0.14273077249526978, + 0.04295956343412399, + 0.10201602429151535 + ], + "credit_map": [ + 0.022488506510853767, + 0.06698691844940186, + 0.05269841477274895, + 0.04211333021521568, + 0.01662842370569706, + 0.015656528994441032, + 0.01749609224498272, + 0.018192410469055176, + 0.01506196428090334, + 0.00870821438729763, + 0.009788578376173973, + 0.010969843715429306, + 0.012082689441740513, + 0.006809191312640905, + 0.009611603803932667, + 0.007096676155924797, + 0.019532667472958565, + 0.012661894783377647, + 0.014958695508539677, + 0.015726417303085327, + 0.01953514851629734, + 0.012594164349138737, + 0.018178816884756088, + 0.025469748303294182, + 0.009075080044567585, + 0.006329812575131655, + 0.012919970788061619, + 0.014108884148299694, + 0.007680126931518316, + 0.010840124450623989, + 0.010754205286502838, + 0.015370089560747147, + 0.0033898460678756237, + 0.004443065728992224, + 0.004556513857096434, + 0.005054090172052383, + 0.007004315033555031, + 0.0048095230013132095, + 0.0033161030150949955, + 0.003894834080711007, + 0.005295282229781151, + 0.004643816500902176, + 0.00646210229024291, + 0.005243154242634773, + 0.008027573116123676, + 0.004494674038141966, + 0.0034129973500967026, + 0.0057801674120128155, + 0.008022822439670563, + 0.01269838958978653, + 0.005607489962130785, + 0.007021780591458082, + 0.013190762139856815, + 0.010859441012144089, + 0.008618908934295177, + 0.008320721797645092, + 0.021043622866272926, + 0.024948956444859505, + 0.013611428439617157, + 0.01324445754289627, + 0.02056180126965046, + 0.021533258259296417, + 0.01780407689511776, + 0.009983167983591557, + 0.01028033159673214, + 0.006335955113172531, + 0.0036516066174954176, + 0.004094819538295269, + 0.0049849399365484715, + 0.005482736509293318, + 0.003969261888414621, + 0.004159913863986731, + 0.011530612595379353, + 0.011938660405576229, + 0.006911196745932102, + 0.007706756237894297, + 0.02450534887611866, + 0.010068868286907673, + 0.00905892625451088, + 0.02029566653072834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.011299780569970608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046506100334227085 + ], + "opposite_logprob_tokens": [ + -5.379650592803955 + ], + "logprob_reward": -0.0046506100334227085, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046506100334227085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.224856436252594, + 0.2361634224653244, + 0.08647503703832626, + 0.08726406842470169, + 0.06979270279407501, + 0.029963422566652298, + 0.03415596857666969, + 0.061324622482061386, + 0.04173459857702255, + 0.12826970219612122 + ], + "credit_map": [ + 0.020036669448018074, + 0.030681367963552475, + 0.03996487334370613, + 0.0542585551738739, + 0.023194218054413795, + 0.017172563821077347, + 0.01695864275097847, + 0.02258954755961895, + 0.06640764325857162, + 0.027019290253520012, + 0.040611013770103455, + 0.02835824154317379, + 0.021056856960058212, + 0.015411988832056522, + 0.02204575389623642, + 0.015252647921442986, + 0.01298221293836832, + 0.007349927444010973, + 0.010512135922908783, + 0.00970639381557703, + 0.009581644088029861, + 0.005858307238668203, + 0.01945122703909874, + 0.0110331941395998, + 0.008969712071120739, + 0.005569780711084604, + 0.0077446987852454185, + 0.021023036912083626, + 0.006529987324029207, + 0.010257178917527199, + 0.00877836998552084, + 0.018391303718090057, + 0.006003209389746189, + 0.012521376833319664, + 0.020165761932730675, + 0.00461457809433341, + 0.007869154214859009, + 0.006738428957760334, + 0.0035944958217442036, + 0.00828570406883955, + 0.003981640562415123, + 0.0033173353876918554, + 0.002629881491884589, + 0.002903882646933198, + 0.0052722604013979435, + 0.004630635026842356, + 0.0028931773267686367, + 0.004334612749516964, + 0.003281412646174431, + 0.005855869967490435, + 0.003476752433925867, + 0.0030037080869078636, + 0.004192260093986988, + 0.003698547137901187, + 0.0040953755378723145, + 0.0065520452335476875, + 0.007535419426858425, + 0.01286225114017725, + 0.004601799882948399, + 0.005256504286080599, + 0.007987748831510544, + 0.010237145237624645, + 0.008327309973537922, + 0.004516445565968752, + 0.007714559789747, + 0.0074096485041081905, + 0.004185976926237345, + 0.0037295278161764145, + 0.005086986813694239, + 0.0057114968076348305, + 0.003907348029315472, + 0.003989059943705797, + 0.012011716142296791, + 0.015337934717535973, + 0.007784242741763592, + 0.007785097695887089, + 0.04147521033883095, + 0.012719438411295414, + 0.009693322703242302, + 0.02146274968981743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0046506100334227085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648676797747612 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.008648676797747612, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648676797747612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1856531947851181, + 0.05795946717262268, + 0.11495980620384216, + 0.07133332639932632, + 0.04407297074794769, + 0.07634548842906952, + 0.039854053407907486, + 0.10724500566720963, + 0.07858065515756607, + 0.22399604320526123 + ], + "credit_map": [ + 0.01711844466626644, + 0.044088561087846756, + 0.02944670058786869, + 0.03720471262931824, + 0.016567664220929146, + 0.013985292986035347, + 0.01502301450818777, + 0.012218805961310863, + 0.011811433359980583, + 0.00778307905420661, + 0.006751922890543938, + 0.008403530344367027, + 0.007395339198410511, + 0.004906558897346258, + 0.006424427032470703, + 0.004483172670006752, + 0.015946464613080025, + 0.011718395166099072, + 0.013172880746424198, + 0.012795986607670784, + 0.01649473048746586, + 0.009634758345782757, + 0.016458962112665176, + 0.01873761974275112, + 0.0076601626351475716, + 0.00508172158151865, + 0.007597656920552254, + 0.013566195964813232, + 0.007258749566972256, + 0.00828546192497015, + 0.009575532749295235, + 0.012307842262089252, + 0.005400550086051226, + 0.005037122406065464, + 0.005627427250146866, + 0.004941050428897142, + 0.006199217867106199, + 0.006013990845531225, + 0.005144675727933645, + 0.005708934273570776, + 0.007660197094082832, + 0.008698615245521069, + 0.009055042639374733, + 0.006360467988997698, + 0.01436296384781599, + 0.013513204641640186, + 0.005855147261172533, + 0.01083985436707735, + 0.004411433823406696, + 0.006574342492967844, + 0.0037272197660058737, + 0.0041847508400678635, + 0.00637135561555624, + 0.004903891123831272, + 0.00532224727794528, + 0.004358813166618347, + 0.015078610740602016, + 0.01972186751663685, + 0.010509157553315163, + 0.01165684126317501, + 0.010074282996356487, + 0.00903110858052969, + 0.020479410886764526, + 0.010693715885281563, + 0.012736688368022442, + 0.01054196897894144, + 0.007318584248423576, + 0.009231207892298698, + 0.0071913390420377254, + 0.011302067898213863, + 0.011938934214413166, + 0.008319864980876446, + 0.05193263664841652, + 0.01765328086912632, + 0.010864497162401676, + 0.009824950248003006, + 0.0818784087896347, + 0.022132543846964836, + 0.011615892872214317, + 0.018093829974532127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.008648676797747612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270424872636795 + ], + "opposite_logprob_tokens": [ + -2.127042531967163 + ], + "logprob_reward": -0.1270424872636795, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270424872636795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16890062391757965, + 0.09244225919246674, + 0.08100280910730362, + 0.23605535924434662, + 0.04048730432987213, + 0.05192391946911812, + 0.04263921082019806, + 0.0826806053519249, + 0.10260605812072754, + 0.10126190632581711 + ], + "credit_map": [ + 0.01579860970377922, + 0.04028915986418724, + 0.031121673062443733, + 0.035249482840299606, + 0.013339516706764698, + 0.011569803580641747, + 0.010838707908987999, + 0.010693680495023727, + 0.012485126033425331, + 0.013445901684463024, + 0.009910451248288155, + 0.014088192023336887, + 0.01391612458974123, + 0.009021736681461334, + 0.012454204261302948, + 0.007120524998754263, + 0.011201976798474789, + 0.008650075644254684, + 0.008249073289334774, + 0.008674818091094494, + 0.012538953684270382, + 0.006904062815010548, + 0.014675471000373363, + 0.010108379647135735, + 0.025178270414471626, + 0.014930780977010727, + 0.014147312380373478, + 0.06294483691453934, + 0.01732710935175419, + 0.02833499386906624, + 0.03035752661526203, + 0.04283451661467552, + 0.00692030182108283, + 0.005796909797936678, + 0.004481072537600994, + 0.004501678980886936, + 0.005815562792122364, + 0.005593485198915005, + 0.003499335143715143, + 0.0038789589889347553, + 0.005439209286123514, + 0.005788184702396393, + 0.005662835668772459, + 0.0044801621697843075, + 0.011806577444076538, + 0.006463340949267149, + 0.005127375945448875, + 0.007156237028539181, + 0.0033409029711037874, + 0.0065726833418011665, + 0.004338240250945091, + 0.004000648856163025, + 0.006996851414442062, + 0.005211136769503355, + 0.004565132781863213, + 0.0076136100105941296, + 0.008253172971308231, + 0.015670394524931908, + 0.007589211221784353, + 0.007894326001405716, + 0.010188012383878231, + 0.007174869999289513, + 0.017345955595374107, + 0.008564657531678677, + 0.013257347047328949, + 0.016828421503305435, + 0.008667600341141224, + 0.008377845399081707, + 0.014833712950348854, + 0.013768056407570839, + 0.016489336267113686, + 0.010383729822933674, + 0.021194852888584137, + 0.009095977991819382, + 0.00517737353220582, + 0.004773213993757963, + 0.035063572227954865, + 0.011338390409946442, + 0.005919830873608589, + 0.00869869627058506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.1270424872636795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495576187968254 + ], + "opposite_logprob_tokens": [ + -2.9299557209014893 + ], + "logprob_reward": -0.05495576187968254, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495576187968254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2348913550376892, + 0.059773121029138565, + 0.07536382228136063, + 0.13575221598148346, + 0.17758026719093323, + 0.061758652329444885, + 0.05595545843243599, + 0.07963746041059494, + 0.04049305617809296, + 0.07879465818405151 + ], + "credit_map": [ + 0.018309513106942177, + 0.04311322048306465, + 0.04306793957948685, + 0.08098999410867691, + 0.012779494747519493, + 0.009663380682468414, + 0.01310410350561142, + 0.013863686472177505, + 0.010620173066854477, + 0.00816576462239027, + 0.007654762826859951, + 0.006951652467250824, + 0.007779695559293032, + 0.005677633918821812, + 0.007123593240976334, + 0.005799842067062855, + 0.012319471687078476, + 0.006622807588428259, + 0.015580615028738976, + 0.0064478483982384205, + 0.0055541410110890865, + 0.004630488343536854, + 0.01297564897686243, + 0.011232784017920494, + 0.015422987751662731, + 0.009890034794807434, + 0.010439463891088963, + 0.042654190212488174, + 0.012345951981842518, + 0.017358070239424706, + 0.012041090987622738, + 0.015600417740643024, + 0.009951711632311344, + 0.0347304493188858, + 0.056328047066926956, + 0.014575550332665443, + 0.01522913109511137, + 0.015260335989296436, + 0.008860216476023197, + 0.022644801065325737, + 0.006700028199702501, + 0.00743793323636055, + 0.009982415474951267, + 0.009632252156734467, + 0.009710773825645447, + 0.005665395874530077, + 0.007173408288508654, + 0.0054564387537539005, + 0.0073312618769705296, + 0.008549967780709267, + 0.006202467251569033, + 0.006546856369823217, + 0.008417571894824505, + 0.008442570455372334, + 0.00540965748950839, + 0.005055100657045841, + 0.009615691378712654, + 0.015460455790162086, + 0.008846348151564598, + 0.008627483621239662, + 0.009505102410912514, + 0.006274071056395769, + 0.013140913099050522, + 0.008167391642928123, + 0.008081325329840183, + 0.005887583829462528, + 0.0033451588824391365, + 0.0031079878099262714, + 0.004196717403829098, + 0.006683731451630592, + 0.0047390880063176155, + 0.004451458342373371, + 0.01736305095255375, + 0.008542431518435478, + 0.003885069163516164, + 0.0051703364588320255, + 0.024478822946548462, + 0.010020808316767216, + 0.004205981735140085, + 0.005128142423927784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.05495576187968254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide-angle shot slowly sweeps across a rain-swept, rain-soaked urban canyon, capturing scenes of catastrophic destruction within a dark, dystopian sci-fi aesthetic. From a turbulent, inky-black sky, massive, crystalline hailstones, some the size of smaller boulders with jagged, faceted surfaces, descend at terrifying speed. They impact with explosive force, violently shattering the ornate, neo-futuristic facades of towering skyscrapers. The camera follows one such impact as a hailstone smashes into a gargoyle-adorned windowsill, sending a shower of black stone and gleaming ice shards into the cracks below. At street level, the sleek, aerodynamic shells of advanced vehicles are systematically crushed and dented by the relentless ice assault. Large, unstable holographic advertisements flicker erratically between buildings, their vibrant cyan and purplish-red light casting chaotic patterns, shifting reflections across the wet asphalt, translucent hailstones, and the smoke rising from short-circuited machinery, creating an atmosphere of overwhelming chaos both terrifying and captivating.", + "seed": 15552, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_7/index_8938_seed_15552_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.050699468795210124, + "vlm_logprob": -0.050699468795210124, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.050699468795210124, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007459169719368219 + ], + "opposite_logprob_tokens": [ + -5.3824591636657715 + ], + "logprob_reward": -0.007459169719368219, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.007459169719368219, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13804319500923157, + 0.07509002089500427, + 0.09327568858861923, + 0.09030180424451828, + 0.1664036363363266, + 0.12620478868484497, + 0.11634550988674164, + 0.052250102162361145, + 0.09978517889976501, + 0.0423000305891037 + ], + "credit_map": [ + 0.008135027252137661, + 0.019608210772275925, + 0.013356884941458702, + 0.02430543303489685, + 0.027482910081744194, + 0.01955566555261612, + 0.012730355374515057, + 0.012868715450167656, + 0.008194707334041595, + 0.010584929026663303, + 0.006417905446141958, + 0.008996550925076008, + 0.015218516811728477, + 0.011763308197259903, + 0.004713284783065319, + 0.009200824424624443, + 0.010846775025129318, + 0.011008985340595245, + 0.016018953174352646, + 0.009263014420866966, + 0.010776127688586712, + 0.007795354817062616, + 0.008274300023913383, + 0.019292179495096207, + 0.01456589624285698, + 0.009619683027267456, + 0.00879590678960085, + 0.009288113564252853, + 0.012258298695087433, + 0.014590233564376831, + 0.009758949279785156, + 0.011424737982451916, + 0.013407791964709759, + 0.04178975522518158, + 0.017819605767726898, + 0.011208620853722095, + 0.012509635649621487, + 0.04628148674964905, + 0.012344244867563248, + 0.011042511090636253, + 0.016858374699950218, + 0.017891140654683113, + 0.016988595947623253, + 0.019908957183361053, + 0.01655540056526661, + 0.011992090381681919, + 0.016799267381429672, + 0.009210966527462006, + 0.010830860584974289, + 0.013126324862241745, + 0.01607496477663517, + 0.023121302947402, + 0.014701713807880878, + 0.008680474944412708, + 0.016873808577656746, + 0.012936064973473549, + 0.008048774674534798, + 0.005479638930410147, + 0.008424191735684872, + 0.0058101569302380085, + 0.006775617133826017, + 0.006843770854175091, + 0.005539737641811371, + 0.005328216124325991, + 0.011549466289579868, + 0.010079571045935154, + 0.012342745438218117, + 0.012975669465959072, + 0.01408236101269722, + 0.01254512369632721, + 0.013271787203848362, + 0.012938463129103184, + 0.0070313094183802605, + 0.004704555030912161, + 0.0038720243610441685, + 0.004511181265115738, + 0.005821190774440765, + 0.005403556860983372, + 0.004884696565568447, + 0.006071519572287798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.007459169719368219 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556307949125767 + ], + "opposite_logprob_tokens": [ + -4.387556076049805 + ], + "logprob_reward": -0.012556307949125767, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556307949125767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16460153460502625, + 0.14096936583518982, + 0.16049405932426453, + 0.1646650731563568, + 0.09913065284490585, + 0.07681603729724884, + 0.05694343522191048, + 0.05365603044629097, + 0.05702412873506546, + 0.025699635967612267 + ], + "credit_map": [ + 0.009711421094834805, + 0.020811274647712708, + 0.01592261716723442, + 0.026681717485189438, + 0.035184238106012344, + 0.024563057348132133, + 0.016072522848844528, + 0.015654707327485085, + 0.014993852935731411, + 0.017231378704309464, + 0.010414530523121357, + 0.01592232659459114, + 0.03295767679810524, + 0.021198756992816925, + 0.011345039121806622, + 0.01690581999719143, + 0.011507832445204258, + 0.018404312431812286, + 0.020858319476246834, + 0.016385119408369064, + 0.019579393789172173, + 0.015975983813405037, + 0.012198681011795998, + 0.04558442905545235, + 0.026363329961895943, + 0.01974683627486229, + 0.01284630224108696, + 0.015903634950518608, + 0.022419290617108345, + 0.028032943606376648, + 0.018575523048639297, + 0.020777229219675064, + 0.014443245716392994, + 0.012547863647341728, + 0.014929031021893024, + 0.011593025177717209, + 0.01020207442343235, + 0.011371230706572533, + 0.0136887077242136, + 0.010355480946600437, + 0.012224472127854824, + 0.011643024161458015, + 0.008637133985757828, + 0.009509027004241943, + 0.010635803453624249, + 0.007759304717183113, + 0.009381117299199104, + 0.007026157807558775, + 0.0051007927395403385, + 0.006778378505259752, + 0.008061710745096207, + 0.008476044051349163, + 0.006592113524675369, + 0.00913840252906084, + 0.0068235634826123714, + 0.005972431972622871, + 0.006577422376722097, + 0.005036395508795977, + 0.007097206544131041, + 0.006711823400110006, + 0.007588723674416542, + 0.007242949213832617, + 0.005856848321855068, + 0.007544662803411484, + 0.006074181292206049, + 0.006030607502907515, + 0.005903979763388634, + 0.00605053873732686, + 0.006228357087820768, + 0.004999977070838213, + 0.010011277161538601, + 0.011725211516022682, + 0.0040806857869029045, + 0.0030240844935178757, + 0.002406222512945533, + 0.00307585671544075, + 0.0026517431251704693, + 0.0022718352265655994, + 0.003915862180292606, + 0.004273346159607172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.012556307949125767 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083466000854969 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.011083466000854969, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083466000854969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09916119277477264, + 0.06170452758669853, + 0.10021704435348511, + 0.1253596693277359, + 0.18164324760437012, + 0.11959131062030792, + 0.08474578708410263, + 0.11068302392959595, + 0.0877363309264183, + 0.0291578508913517 + ], + "credit_map": [ + 0.00541171059012413, + 0.012349630706012249, + 0.010996759869158268, + 0.017866631969809532, + 0.018684566020965576, + 0.013445315882563591, + 0.011478332802653313, + 0.008928242139518261, + 0.006477572955191135, + 0.007076992187649012, + 0.0049039884470403194, + 0.006732537876814604, + 0.015440000221133232, + 0.00882750004529953, + 0.004766746889799833, + 0.007479188963770866, + 0.007706409320235252, + 0.009158596396446228, + 0.020627891644835472, + 0.008867735974490643, + 0.01104517001658678, + 0.010061497800052166, + 0.004715486895292997, + 0.028034253045916557, + 0.018149573355913162, + 0.01260636281222105, + 0.010023936629295349, + 0.011660726740956306, + 0.015377267263829708, + 0.021185070276260376, + 0.019417747855186462, + 0.016938986256718636, + 0.013920878060162067, + 0.0437125638127327, + 0.01124515663832426, + 0.015409477986395359, + 0.01734897680580616, + 0.04956277832388878, + 0.0173470638692379, + 0.013096368871629238, + 0.015597386285662651, + 0.01433255523443222, + 0.014619303867220879, + 0.019081033766269684, + 0.01435401663184166, + 0.009621188044548035, + 0.02185109630227089, + 0.010134733282029629, + 0.0072310264222323895, + 0.009838352911174297, + 0.011677234433591366, + 0.015951007604599, + 0.010303972288966179, + 0.009158275090157986, + 0.011685346253216267, + 0.00890057347714901, + 0.014675232581794262, + 0.013722260482609272, + 0.016908014193177223, + 0.013877957127988338, + 0.014518996700644493, + 0.011069752275943756, + 0.015253549441695213, + 0.010657268576323986, + 0.0101347416639328, + 0.009027552790939808, + 0.009302628226578236, + 0.008178889751434326, + 0.008141368627548218, + 0.007224126718938351, + 0.01780969649553299, + 0.0179173294454813, + 0.004714848008006811, + 0.003067504381760955, + 0.0024196146987378597, + 0.0033425260335206985, + 0.004231571219861507, + 0.0030937474220991135, + 0.0041078804060816765, + 0.004180159419775009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.011083466000854969 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035246849060059 + ], + "opposite_logprob_tokens": [ + -1.9103524684906006 + ], + "logprob_reward": -0.16035246849060059, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035246849060059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13324114680290222, + 0.0639088824391365, + 0.12493587285280228, + 0.07344140112400055, + 0.41446858644485474, + 0.05134979262948036, + 0.04239562898874283, + 0.041550010442733765, + 0.038256943225860596, + 0.016451740637421608 + ], + "credit_map": [ + 0.00757752126082778, + 0.015625005587935448, + 0.011732067912817001, + 0.017072761431336403, + 0.030273426324129105, + 0.020713066682219505, + 0.014285258948802948, + 0.01596204750239849, + 0.006134653463959694, + 0.006005776114761829, + 0.004564844071865082, + 0.009808332659304142, + 0.010668766684830189, + 0.007796465884894133, + 0.010615588165819645, + 0.008314456790685654, + 0.008588057942688465, + 0.012582571245729923, + 0.024187356233596802, + 0.010961109772324562, + 0.012552673928439617, + 0.013772932812571526, + 0.006604603957384825, + 0.03568657860159874, + 0.012478434480726719, + 0.00559663400053978, + 0.005365071818232536, + 0.004177568480372429, + 0.007293466944247484, + 0.007297233212739229, + 0.022454481571912766, + 0.008778505027294159, + 0.02778743766248226, + 0.11360450834035873, + 0.024906635284423828, + 0.027143191546201706, + 0.02883273735642433, + 0.13913212716579437, + 0.027770057320594788, + 0.02529188245534897, + 0.008703220635652542, + 0.006177577190101147, + 0.00555916503071785, + 0.006464590784162283, + 0.006500220857560635, + 0.005644432734698057, + 0.0072880154475569725, + 0.005012566689401865, + 0.004914405290037394, + 0.005541657097637653, + 0.004262984264642, + 0.00569876516237855, + 0.005534895695745945, + 0.005010797642171383, + 0.005890360567718744, + 0.005541764199733734, + 0.005197850987315178, + 0.004665678832679987, + 0.006083695683628321, + 0.004562170244753361, + 0.006227666512131691, + 0.004838025663048029, + 0.005353041458874941, + 0.004621880128979683, + 0.005130006931722164, + 0.004217374604195356, + 0.0039461879059672356, + 0.004722832702100277, + 0.004985201172530651, + 0.003952771425247192, + 0.004911383613944054, + 0.006391183938831091, + 0.0020826850086450577, + 0.0021407571621239185, + 0.0015310212038457394, + 0.001465245964936912, + 0.002765234559774399, + 0.0020610084757208824, + 0.0019933218136429787, + 0.002412466099485755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.16035246849060059 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062045931816101074 + ], + "opposite_logprob_tokens": [ + -2.8120460510253906 + ], + "logprob_reward": -0.062045931816101074, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062045931816101074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10099761933088303, + 0.055973589420318604, + 0.1396338939666748, + 0.07658654451370239, + 0.07915903627872467, + 0.08059610426425934, + 0.18448644876480103, + 0.11497396975755692, + 0.1390947550535202, + 0.028498033061623573 + ], + "credit_map": [ + 0.008355467580258846, + 0.013404280878603458, + 0.010456638410687447, + 0.016309941187500954, + 0.014378953725099564, + 0.014815513044595718, + 0.01074721198529005, + 0.012529613450169563, + 0.00626537436619401, + 0.00674333143979311, + 0.004629255272448063, + 0.008460426703095436, + 0.008278382010757923, + 0.008556830696761608, + 0.005879961419850588, + 0.007160032168030739, + 0.012818957678973675, + 0.014196839183568954, + 0.024161551147699356, + 0.013255522586405277, + 0.015436604619026184, + 0.015298722311854362, + 0.008550341241061687, + 0.03591535985469818, + 0.012120399624109268, + 0.005682306829839945, + 0.009658890776336193, + 0.005190159194171429, + 0.009033069014549255, + 0.005187796428799629, + 0.023789577186107635, + 0.0059243422001600266, + 0.010673198848962784, + 0.00940609723329544, + 0.007504160050302744, + 0.008143927901983261, + 0.014356127940118313, + 0.010202594101428986, + 0.011324545368552208, + 0.007548391819000244, + 0.01189237367361784, + 0.009273182600736618, + 0.008521003648638725, + 0.011574077419936657, + 0.010888945311307907, + 0.008672347292304039, + 0.012357768602669239, + 0.0074164001271128654, + 0.019917963072657585, + 0.022920716553926468, + 0.027375150471925735, + 0.038159217685461044, + 0.0234904196113348, + 0.018113505095243454, + 0.01965673826634884, + 0.014852747321128845, + 0.013591180555522442, + 0.01248131413012743, + 0.016020195558667183, + 0.018063129857182503, + 0.016913581639528275, + 0.0094067407771945, + 0.012826040387153625, + 0.01567179523408413, + 0.01322094164788723, + 0.015041218139231205, + 0.013264982029795647, + 0.016115443781018257, + 0.02056347392499447, + 0.01894584484398365, + 0.019402705132961273, + 0.022540142759680748, + 0.005322565790265799, + 0.00358304800465703, + 0.00329454499296844, + 0.002191305859014392, + 0.00472769932821393, + 0.0027945011388510466, + 0.0030836095102131367, + 0.003500758670270443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.062045931816101074 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide-angle shot slowly sweeps across a rain-swept, rain-soaked urban canyon, capturing scenes of catastrophic destruction within a dark, dystopian sci-fi aesthetic. From a turbulent, inky-black sky, massive, crystalline hailstones, some the size of smaller boulders with jagged, faceted surfaces, descend at terrifying speed. They impact with explosive force, violently shattering the ornate, neo-futuristic facades of towering skyscrapers. The camera follows one such impact as a hailstone smashes into a gargoyle-adorned windowsill, sending a shower of black stone and gleaming ice shards into the cracks below. At street level, the sleek, aerodynamic shells of advanced vehicles are systematically crushed and dented by the relentless ice assault. Large, unstable holographic advertisements flicker erratically between buildings, their vibrant cyan and purplish-red light casting chaotic patterns, shifting reflections across the wet asphalt, translucent hailstones, and the smoke rising from short-circuited machinery, creating an atmosphere of overwhelming chaos both terrifying and captivating.", + "seed": 15553, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_7/index_8938_seed_15553_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05797806205227971, + "vlm_logprob": -0.05797806205227971, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.05797806205227971, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01357231941074133 + ], + "opposite_logprob_tokens": [ + -4.5135722160339355 + ], + "logprob_reward": -0.01357231941074133, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01357231941074133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26243895292282104, + 0.10907284170389175, + 0.03662895783782005, + 0.059285469353199005, + 0.02608633041381836, + 0.06058254837989807, + 0.17435425519943237, + 0.13401365280151367, + 0.06558757275342941, + 0.07194938510656357 + ], + "credit_map": [ + 0.04007410630583763, + 0.02868795208632946, + 0.018480127677321434, + 0.02099955826997757, + 0.03515194356441498, + 0.029520807787775993, + 0.02746068499982357, + 0.0620637871325016, + 0.013725589029490948, + 0.016283174976706505, + 0.009685767814517021, + 0.008370038121938705, + 0.014736423268914223, + 0.015093650668859482, + 0.011914622038602829, + 0.019263580441474915, + 0.0037777021061629057, + 0.004434502683579922, + 0.004712843801826239, + 0.003964823670685291, + 0.00494754035025835, + 0.006211861502379179, + 0.003651805454865098, + 0.00492787966504693, + 0.006712191738188267, + 0.010364213958382607, + 0.011005988344550133, + 0.0058443318121135235, + 0.005132442805916071, + 0.00826035626232624, + 0.005040392279624939, + 0.006925553549081087, + 0.0033972689416259527, + 0.003931236919015646, + 0.003753031836822629, + 0.003779488615691662, + 0.0020711631514132023, + 0.002475177403539419, + 0.0036234913859516382, + 0.003055473556742072, + 0.013708223588764668, + 0.0067023723386228085, + 0.007077194284647703, + 0.006787243764847517, + 0.005573555361479521, + 0.006011151243001223, + 0.006916483864188194, + 0.007806322071701288, + 0.017918327823281288, + 0.015679607167840004, + 0.0273133572191, + 0.026417065411806107, + 0.01921176351606846, + 0.02474047988653183, + 0.0272971261292696, + 0.01577652618288994, + 0.01213896181434393, + 0.009528189897537231, + 0.01705062948167324, + 0.01837487518787384, + 0.014537112787365913, + 0.01556189265102148, + 0.014860577881336212, + 0.03196141496300697, + 0.008732098154723644, + 0.00844163354486227, + 0.005956312641501427, + 0.00876628514379263, + 0.01202851440757513, + 0.008416787721216679, + 0.00502377562224865, + 0.008222165517508984, + 0.013230123557150364, + 0.006505392957478762, + 0.006415440700948238, + 0.009679854847490788, + 0.008273283019661903, + 0.005982525646686554, + 0.01035003736615181, + 0.011512719094753265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.01357231941074133 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748977117240429 + ], + "opposite_logprob_tokens": [ + -5.006749153137207 + ], + "logprob_reward": -0.006748977117240429, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748977117240429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2558331787586212, + 0.11927847564220428, + 0.05602617934346199, + 0.07540803402662277, + 0.05760853365063667, + 0.04832044988870621, + 0.16703182458877563, + 0.1138174906373024, + 0.0617256723344326, + 0.044950179755687714 + ], + "credit_map": [ + 0.046973854303359985, + 0.020067123696208, + 0.015817683190107346, + 0.01554903108626604, + 0.021703744307160378, + 0.02863621711730957, + 0.02959856390953064, + 0.07748694717884064, + 0.01551870908588171, + 0.017682312056422234, + 0.010414117015898228, + 0.009419958107173443, + 0.014690284617245197, + 0.016478758305311203, + 0.013462762348353863, + 0.02161157876253128, + 0.006609953474253416, + 0.009454741142690182, + 0.005920163821429014, + 0.004823305178433657, + 0.008576749823987484, + 0.008523164317011833, + 0.005254119634628296, + 0.006863979622721672, + 0.010212785564363003, + 0.013614367693662643, + 0.01479067001491785, + 0.00679378816857934, + 0.00609510438516736, + 0.009404699318110943, + 0.006139704957604408, + 0.008356910198926926, + 0.011660926975309849, + 0.007307369261980057, + 0.008166293613612652, + 0.007990681566298008, + 0.004487011581659317, + 0.005795774981379509, + 0.004866532050073147, + 0.007333945948630571, + 0.013058776967227459, + 0.005309863947331905, + 0.00617971969768405, + 0.005221985280513763, + 0.003674754174426198, + 0.0036201716866344213, + 0.004927962552756071, + 0.006327212322503328, + 0.01463975477963686, + 0.014292259700596333, + 0.024556880816817284, + 0.02518482878804207, + 0.01735372096300125, + 0.023427924141287804, + 0.030018024146556854, + 0.017558421939611435, + 0.013531383126974106, + 0.006086174864321947, + 0.008540792390704155, + 0.022424623370170593, + 0.013890336267650127, + 0.012763897888362408, + 0.01482127234339714, + 0.02175901271402836, + 0.007541946601122618, + 0.007313656620681286, + 0.0060150232166051865, + 0.008717401884496212, + 0.010792070999741554, + 0.008015279658138752, + 0.005326281767338514, + 0.00800400786101818, + 0.00993321742862463, + 0.004290983080863953, + 0.004597937688231468, + 0.004417356103658676, + 0.004841858055442572, + 0.005177988670766354, + 0.005313174333423376, + 0.006377665791660547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.006748977117240429 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005957346875220537 + ], + "opposite_logprob_tokens": [ + -5.130957126617432 + ], + "logprob_reward": -0.005957346875220537, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005957346875220537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15440337359905243, + 0.07260055094957352, + 0.04222015663981438, + 0.06849482655525208, + 0.018930617719888687, + 0.0766303688287735, + 0.3398151099681854, + 0.09663748741149902, + 0.06911643594503403, + 0.06115111708641052 + ], + "credit_map": [ + 0.02583215944468975, + 0.014176193624734879, + 0.011547260917723179, + 0.011114880442619324, + 0.015629954636096954, + 0.017932510003447533, + 0.01762363314628601, + 0.04054677113890648, + 0.009748958982527256, + 0.011977015994489193, + 0.0062465607188642025, + 0.005273347720503807, + 0.009379149414598942, + 0.009992975741624832, + 0.007276422344148159, + 0.012706112116575241, + 0.004919294733554125, + 0.006027125753462315, + 0.004482214339077473, + 0.0032883889507502317, + 0.007785812485963106, + 0.006419018376618624, + 0.0037570183631032705, + 0.005541279446333647, + 0.010199510492384434, + 0.010662269778549671, + 0.009461754932999611, + 0.007751932833343744, + 0.005602541379630566, + 0.01017573568969965, + 0.007537573575973511, + 0.007103500422090292, + 0.0028412993997335434, + 0.002178905066102743, + 0.0031015563290566206, + 0.002257231855764985, + 0.0011002338724210858, + 0.0031446360517293215, + 0.002042884938418865, + 0.0022638672962784767, + 0.014412603341042995, + 0.009807081893086433, + 0.008302368223667145, + 0.007519861683249474, + 0.008252698928117752, + 0.010105028748512268, + 0.008718621917068958, + 0.009512102231383324, + 0.023652851581573486, + 0.02716204896569252, + 0.05287682265043259, + 0.04736804962158203, + 0.03592456132173538, + 0.05105336010456085, + 0.06298133730888367, + 0.03879605233669281, + 0.014502857811748981, + 0.0055425879545509815, + 0.008197633549571037, + 0.019855206832289696, + 0.012980982661247253, + 0.011905670166015625, + 0.011696314439177513, + 0.011956224218010902, + 0.00896948017179966, + 0.008589019067585468, + 0.008282090537250042, + 0.010815778747200966, + 0.010464494116604328, + 0.008492688648402691, + 0.006466869730502367, + 0.007036014460027218, + 0.009711869060993195, + 0.005714606028050184, + 0.00793803296983242, + 0.009287686087191105, + 0.006742897909134626, + 0.004860027693212032, + 0.008539954200387001, + 0.008356036618351936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.005957346875220537 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015688717365265 + ], + "opposite_logprob_tokens": [ + -1.701568841934204 + ], + "logprob_reward": -0.2015688717365265, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015688717365265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2697013020515442, + 0.06616675853729248, + 0.0303630568087101, + 0.059977203607559204, + 0.02177383191883564, + 0.10204046219587326, + 0.11083020269870758, + 0.14314202964305878, + 0.12829525768756866, + 0.06770991533994675 + ], + "credit_map": [ + 0.05295445770025253, + 0.02231522649526596, + 0.01601005718111992, + 0.015853367745876312, + 0.024302951991558075, + 0.024953728541731834, + 0.027518255636096, + 0.08579327911138535, + 0.00900629349052906, + 0.010716342367231846, + 0.006093507632613182, + 0.004728481639176607, + 0.008783979341387749, + 0.008959702216088772, + 0.007609815336763859, + 0.010268633253872395, + 0.00354380183853209, + 0.004105524159967899, + 0.0026385837700217962, + 0.0027683707885444164, + 0.0053136409260332584, + 0.004823528230190277, + 0.0026355900336056948, + 0.004534018225967884, + 0.007467225659638643, + 0.008578796871006489, + 0.011237997561693192, + 0.0066836681216955185, + 0.005578207783401012, + 0.008525937795639038, + 0.004516701214015484, + 0.007388669531792402, + 0.0023833890445530415, + 0.0036006146110594273, + 0.002430432476103306, + 0.002888620365411043, + 0.0024778475053608418, + 0.0020727182272821665, + 0.0029231628868728876, + 0.0029970475006848574, + 0.030031876638531685, + 0.011170588433742523, + 0.012943168170750141, + 0.010018942877650261, + 0.006975270342081785, + 0.006569748278707266, + 0.010435475967824459, + 0.013895398937165737, + 0.013148759491741657, + 0.01188584789633751, + 0.014810601249337196, + 0.01746412366628647, + 0.012159008532762527, + 0.01331594679504633, + 0.018658552318811417, + 0.009387357160449028, + 0.014385679736733437, + 0.007356860209256411, + 0.010327835567295551, + 0.02576434798538685, + 0.016515949741005898, + 0.013393584638834, + 0.020315993577241898, + 0.03508177772164345, + 0.017569318413734436, + 0.015970539301633835, + 0.016747618094086647, + 0.021436896175146103, + 0.016877533867955208, + 0.016507785767316818, + 0.011091078631579876, + 0.012094492092728615, + 0.01648605614900589, + 0.0071154385805130005, + 0.007814256474375725, + 0.0057220133021473885, + 0.0068543399684131145, + 0.00794646330177784, + 0.005933527834713459, + 0.009837815538048744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.2015688717365265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204279512166977 + ], + "opposite_logprob_tokens": [ + -2.812042713165283 + ], + "logprob_reward": -0.06204279512166977, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204279512166977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3382957875728607, + 0.0540383905172348, + 0.0411771796643734, + 0.13084407150745392, + 0.041265830397605896, + 0.055885884910821915, + 0.07220133394002914, + 0.10894709825515747, + 0.0807892233133316, + 0.07655522972345352 + ], + "credit_map": [ + 0.06898073852062225, + 0.024942481890320778, + 0.020063843578100204, + 0.01849864423274994, + 0.029619727283716202, + 0.034609969705343246, + 0.03355344757437706, + 0.10802695155143738, + 0.006321372464299202, + 0.009945841506123543, + 0.004611731041222811, + 0.003588286926969886, + 0.006437703501433134, + 0.008326075039803982, + 0.005996912717819214, + 0.00881047360599041, + 0.0045651039108633995, + 0.006290162913501263, + 0.004868280608206987, + 0.00436769425868988, + 0.00549611821770668, + 0.004266335628926754, + 0.00422699423506856, + 0.007096492685377598, + 0.015468723140656948, + 0.02351224608719349, + 0.021601121872663498, + 0.013145222328603268, + 0.01354362815618515, + 0.02176091820001602, + 0.010584102012217045, + 0.011228111572563648, + 0.009040421806275845, + 0.005413560196757317, + 0.005884729791432619, + 0.005119096953421831, + 0.0030416217632591724, + 0.004583878442645073, + 0.0035488842986524105, + 0.004633641801774502, + 0.012679186649620533, + 0.0060813394375145435, + 0.0057625374756753445, + 0.005242213606834412, + 0.006059807725250721, + 0.005474831443279982, + 0.007042078301310539, + 0.007543893530964851, + 0.008343357592821121, + 0.011294415220618248, + 0.008276958018541336, + 0.011063594371080399, + 0.00849186908453703, + 0.010301526635885239, + 0.009055346250534058, + 0.005374279338866472, + 0.018893130123615265, + 0.006772373802959919, + 0.016505777835845947, + 0.011018279939889908, + 0.013427859172224998, + 0.011228017508983612, + 0.012982324697077274, + 0.018119342625141144, + 0.009864079765975475, + 0.011274898424744606, + 0.00975332222878933, + 0.012604592368006706, + 0.010671957395970821, + 0.009987248107790947, + 0.008655130863189697, + 0.007978007197380066, + 0.01291397400200367, + 0.006844427902251482, + 0.01039651408791542, + 0.010865660384297371, + 0.00924707017838955, + 0.005594352260231972, + 0.008668428286910057, + 0.012024813331663609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.06204279512166977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide-angle shot slowly sweeps across a rain-swept, rain-soaked urban canyon, capturing scenes of catastrophic destruction within a dark, dystopian sci-fi aesthetic. From a turbulent, inky-black sky, massive, crystalline hailstones, some the size of smaller boulders with jagged, faceted surfaces, descend at terrifying speed. They impact with explosive force, violently shattering the ornate, neo-futuristic facades of towering skyscrapers. The camera follows one such impact as a hailstone smashes into a gargoyle-adorned windowsill, sending a shower of black stone and gleaming ice shards into the cracks below. At street level, the sleek, aerodynamic shells of advanced vehicles are systematically crushed and dented by the relentless ice assault. Large, unstable holographic advertisements flicker erratically between buildings, their vibrant cyan and purplish-red light casting chaotic patterns, shifting reflections across the wet asphalt, translucent hailstones, and the smoke rising from short-circuited machinery, creating an atmosphere of overwhelming chaos both terrifying and captivating.", + "seed": 15554, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000077/group_7/index_8938_seed_15554_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03047364195808768, + "vlm_logprob": -0.03047364195808768, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.03047364195808768, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007830636575818062 + ], + "opposite_logprob_tokens": [ + -5.757830619812012 + ], + "logprob_reward": -0.007830636575818062, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007830636575818062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11485844850540161, + 0.05859009549021721, + 0.061223164200782776, + 0.11305857449769974, + 0.04834672063589096, + 0.08697236329317093, + 0.06425298750400543, + 0.13005752861499786, + 0.14880330860614777, + 0.17383690178394318 + ], + "credit_map": [ + 0.011082654818892479, + 0.013756155036389828, + 0.0185803584754467, + 0.011830654926598072, + 0.012422674335539341, + 0.016650134697556496, + 0.016225680708885193, + 0.014310126192867756, + 0.0046355524100363255, + 0.012413989752531052, + 0.006925733759999275, + 0.00783655047416687, + 0.006641599349677563, + 0.007815631106495857, + 0.007102194707840681, + 0.005218834150582552, + 0.006003532558679581, + 0.007010101806372404, + 0.013410653918981552, + 0.005518788937479258, + 0.007424044422805309, + 0.008693606592714787, + 0.005215140525251627, + 0.007947293110191822, + 0.011854303069412708, + 0.026806386187672615, + 0.012162361294031143, + 0.015542891807854176, + 0.01599067449569702, + 0.009174791164696217, + 0.012091943062841892, + 0.00943521223962307, + 0.005541491322219372, + 0.009315630421042442, + 0.0063148983754217625, + 0.0045449030585587025, + 0.0047659482806921005, + 0.004204608034342527, + 0.005691507365554571, + 0.00796772912144661, + 0.022061696276068687, + 0.009566528722643852, + 0.014582822099328041, + 0.008876931853592396, + 0.009557581506669521, + 0.006701720878481865, + 0.006878041662275791, + 0.008747031912207603, + 0.006172360386699438, + 0.0107717989012599, + 0.006115666124969721, + 0.009281271137297153, + 0.009373292326927185, + 0.006809561979025602, + 0.00716730672866106, + 0.008561722002923489, + 0.016645153984427452, + 0.023645736277103424, + 0.006775630638003349, + 0.01332926470786333, + 0.023473933339118958, + 0.02392629347741604, + 0.009049669839441776, + 0.013211820274591446, + 0.010371719487011433, + 0.030985035002231598, + 0.01158114057034254, + 0.011574430391192436, + 0.030532876029610634, + 0.028073156252503395, + 0.014609000645577908, + 0.011075939051806927, + 0.01650405488908291, + 0.020687831565737724, + 0.027966322377324104, + 0.015670567750930786, + 0.03050266019999981, + 0.02389005571603775, + 0.027507508173584938, + 0.011107878759503365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.007830636575818062 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046506100334227085 + ], + "opposite_logprob_tokens": [ + -5.379650592803955 + ], + "logprob_reward": -0.0046506100334227085, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046506100334227085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1959637552499771, + 0.06853877007961273, + 0.07828247547149658, + 0.06316540390253067, + 0.06794725358486176, + 0.10240992158651352, + 0.05876421928405762, + 0.16949425637722015, + 0.10023675858974457, + 0.09519720077514648 + ], + "credit_map": [ + 0.014284297823905945, + 0.0513593964278698, + 0.024898773059248924, + 0.015898851677775383, + 0.02684009075164795, + 0.017397623509168625, + 0.021594608202576637, + 0.023690106347203255, + 0.0055542211048305035, + 0.012908711098134518, + 0.007806152570992708, + 0.00915491208434105, + 0.006669068243354559, + 0.007491812575608492, + 0.011961014941334724, + 0.006992881186306477, + 0.006916750222444534, + 0.009844633750617504, + 0.011777468025684357, + 0.007607657462358475, + 0.014080503024160862, + 0.009645644575357437, + 0.0075593669898808, + 0.010850448161363602, + 0.004951507318764925, + 0.01585250161588192, + 0.005580699536949396, + 0.005654058884829283, + 0.010800505056977272, + 0.007997391745448112, + 0.007691290229558945, + 0.0046374499797821045, + 0.007667831145226955, + 0.012368937022984028, + 0.008108871057629585, + 0.007561221718788147, + 0.008781605400145054, + 0.006702899932861328, + 0.007435875944793224, + 0.009320015087723732, + 0.006218630354851484, + 0.006857904605567455, + 0.006420452147722244, + 0.032835453748703, + 0.0068538859486579895, + 0.0038330170791596174, + 0.03263617679476738, + 0.006754408124834299, + 0.0064502377063035965, + 0.011343789286911488, + 0.0065927631221711636, + 0.00875867810100317, + 0.006136727053672075, + 0.01023457758128643, + 0.0041893781162798405, + 0.005058072041720152, + 0.01662229746580124, + 0.0325576514005661, + 0.007528356742113829, + 0.013500351458787918, + 0.040172528475522995, + 0.03944279998540878, + 0.008274015970528126, + 0.011396250687539577, + 0.00811653770506382, + 0.02053685113787651, + 0.008011844009160995, + 0.01018438395112753, + 0.014965499751269817, + 0.01936511881649494, + 0.009749669581651688, + 0.009306848049163818, + 0.009453767910599709, + 0.013159126974642277, + 0.017749840393662453, + 0.01031198725104332, + 0.01092240959405899, + 0.013895539566874504, + 0.012632599100470543, + 0.007071923464536667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0046506100334227085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958176217973232 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.005958176217973232, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958176217973232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0849694088101387, + 0.057271577417850494, + 0.08929049968719482, + 0.0836980938911438, + 0.07049333304166794, + 0.09144198149442673, + 0.09130874276161194, + 0.11659258604049683, + 0.105982705950737, + 0.20895107090473175 + ], + "credit_map": [ + 0.007578316144645214, + 0.007481405511498451, + 0.015555520541965961, + 0.007825818844139576, + 0.008173521608114243, + 0.016535600647330284, + 0.00994956400245428, + 0.011869656853377819, + 0.004374217242002487, + 0.00944492407143116, + 0.007200869265943766, + 0.007283254526555538, + 0.0064202784560620785, + 0.008561188355088234, + 0.007758445106446743, + 0.006228397134691477, + 0.007969175465404987, + 0.013460971415042877, + 0.015476674772799015, + 0.010376228019595146, + 0.011442628689110279, + 0.011474823579192162, + 0.01011693011969328, + 0.008973069489002228, + 0.007304087746888399, + 0.01757238619029522, + 0.007926137186586857, + 0.00910806842148304, + 0.015407753176987171, + 0.010018715634942055, + 0.009738828055560589, + 0.006622116081416607, + 0.006741285789757967, + 0.012693492695689201, + 0.006976453121751547, + 0.008086026646196842, + 0.012948677875101566, + 0.008142641745507717, + 0.005833792965859175, + 0.009070969186723232, + 0.012391791678965092, + 0.0074334596283733845, + 0.009265835396945477, + 0.019754206761717796, + 0.010249449871480465, + 0.007181316148489714, + 0.017642270773649216, + 0.007523653097450733, + 0.008593780919909477, + 0.011854403652250767, + 0.009488762356340885, + 0.013404604978859425, + 0.013605288229882717, + 0.011129584163427353, + 0.009242193773388863, + 0.013990121893584728, + 0.015961093828082085, + 0.02573142945766449, + 0.006381977815181017, + 0.011940407566726208, + 0.0202848669141531, + 0.0210576094686985, + 0.006146380677819252, + 0.009088825434446335, + 0.007216785103082657, + 0.01160699687898159, + 0.01060840580612421, + 0.008644400164484978, + 0.024861708283424377, + 0.023340731859207153, + 0.010561397299170494, + 0.009142270311713219, + 0.018808912485837936, + 0.025263909250497818, + 0.032324790954589844, + 0.02059963345527649, + 0.031280361115932465, + 0.026016654446721077, + 0.03911817818880081, + 0.01553863100707531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.005958176217973232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897776365280151 + ], + "opposite_logprob_tokens": [ + -2.5789778232574463 + ], + "logprob_reward": -0.07897776365280151, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897776365280151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18444034457206726, + 0.061868127435445786, + 0.1472039520740509, + 0.15940307080745697, + 0.07769652456045151, + 0.05640711262822151, + 0.07702585309743881, + 0.11411209404468536, + 0.05211636424064636, + 0.06972657889127731 + ], + "credit_map": [ + 0.013435395434498787, + 0.03902951627969742, + 0.025238867849111557, + 0.0160780418664217, + 0.026521511375904083, + 0.01790936104953289, + 0.02242937870323658, + 0.023798275738954544, + 0.0039902497082948685, + 0.008340387605130672, + 0.008893235586583614, + 0.008580349385738373, + 0.0063340263441205025, + 0.008250411599874496, + 0.00988976750522852, + 0.007589704357087612, + 0.012010996229946613, + 0.01683463156223297, + 0.01994522474706173, + 0.014805045910179615, + 0.025689657777547836, + 0.018807465210556984, + 0.01648278534412384, + 0.022628137841820717, + 0.010012485086917877, + 0.04369943588972092, + 0.010616990737617016, + 0.01430537924170494, + 0.026475822553038597, + 0.02165002003312111, + 0.020653925836086273, + 0.011989006772637367, + 0.007052266504615545, + 0.012244387529790401, + 0.008252249099314213, + 0.007746786344796419, + 0.015136469155550003, + 0.00974207278341055, + 0.0070751081220805645, + 0.010447181761264801, + 0.008791467174887657, + 0.005608123727142811, + 0.01076580211520195, + 0.00603600824251771, + 0.008119114674627781, + 0.006497837603092194, + 0.005055889021605253, + 0.005532870534807444, + 0.010296142660081387, + 0.015507543459534645, + 0.007956995628774166, + 0.011498396284878254, + 0.007958227768540382, + 0.010120357386767864, + 0.00548200448974967, + 0.008206186816096306, + 0.012761489488184452, + 0.033333610743284225, + 0.006238000001758337, + 0.010441028513014317, + 0.01894339546561241, + 0.019012179225683212, + 0.0057732658460736275, + 0.007609122898429632, + 0.0037398652639240026, + 0.012634165585041046, + 0.0038429927080869675, + 0.004918170627206564, + 0.008386124856770039, + 0.009282018058001995, + 0.005205533467233181, + 0.004107494372874498, + 0.007700661197304726, + 0.012218281626701355, + 0.011356424540281296, + 0.007695950102061033, + 0.010383126325905323, + 0.009161654859781265, + 0.006876288913190365, + 0.0043341959826648235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.07897776365280151 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dark, dystopian sci-fi urban canyon with catastrophic destruction?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive, crystalline hailstones the size of smaller boulders descending from a turbulent, inky-black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the hailstones impact with explosive force, shattering neo-futuristic skyscraper facades and crushing advanced vehicle shells?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a specific shot of a hailstone smashing into a gargoyle-adorned windowsill, sending black stone and ice shards into cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a sweeping wide-angle shot that follows the hailstone impact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549510233104229 + ], + "opposite_logprob_tokens": [ + -2.9299509525299072 + ], + "logprob_reward": -0.0549510233104229, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549510233104229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13387584686279297, + 0.07833491265773773, + 0.0777817964553833, + 0.08944538235664368, + 0.07513947784900665, + 0.0808199867606163, + 0.08962178975343704, + 0.1569220870733261, + 0.1340799778699875, + 0.08397874236106873 + ], + "credit_map": [ + 0.010856278240680695, + 0.03466716408729553, + 0.01672251708805561, + 0.010584636591374874, + 0.01541519071906805, + 0.014099792577326298, + 0.016787244006991386, + 0.014743027277290821, + 0.006028010044246912, + 0.011711339466273785, + 0.009398111142218113, + 0.010928843170404434, + 0.010063798166811466, + 0.013147840276360512, + 0.010958279483020306, + 0.006098696496337652, + 0.006056676618754864, + 0.009678669273853302, + 0.009107360616326332, + 0.008386950008571148, + 0.013079585507512093, + 0.01110794860869646, + 0.009612802416086197, + 0.010751801542937756, + 0.006517284084111452, + 0.013085698708891869, + 0.009385963901877403, + 0.009149962104856968, + 0.014813047833740711, + 0.009791682474315166, + 0.017727529630064964, + 0.008974208496510983, + 0.007813602685928345, + 0.011901579797267914, + 0.010124816559255123, + 0.009668429382145405, + 0.007698547560721636, + 0.00988052599132061, + 0.00877902191132307, + 0.009272953495383263, + 0.015534034930169582, + 0.008709804154932499, + 0.011481735855340958, + 0.01546742394566536, + 0.007627585437148809, + 0.006879169959574938, + 0.007923655211925507, + 0.007196572609245777, + 0.009737460874021053, + 0.01096059288829565, + 0.005999127868562937, + 0.01413947343826294, + 0.01932103931903839, + 0.014581112191081047, + 0.008234554901719093, + 0.006648429203778505, + 0.014823473989963531, + 0.023223359137773514, + 0.0051441071555018425, + 0.012821297161281109, + 0.0455440953373909, + 0.04202023521065712, + 0.005556692369282246, + 0.007788816001266241, + 0.011259419843554497, + 0.039191246032714844, + 0.009004511870443821, + 0.010320167988538742, + 0.02125817909836769, + 0.02388894185423851, + 0.01047442015260458, + 0.008683080784976482, + 0.013389688916504383, + 0.012661442160606384, + 0.010183422826230526, + 0.008442440070211887, + 0.012989806942641735, + 0.010237768292427063, + 0.010245892219245434, + 0.005828287452459335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.0549510233104229 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_12/index_tensor([5266, 5266])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_12/index_tensor([5266, 5266])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..13673f4d5f2ba43bca65f4e376e899e5f35e3fa4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_12/index_tensor([5266, 5266])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This static, tight, eye-level shot, rendered with detailed, grounded realism, focuses on the determined face of an Asian man in his twenties playing pool, leaning over the table to line up a key shot. The camera captures subtle details: the subtle wrinkle of his brow beneath his neatly cropped hair, and the intensely focused reflection of a single, low-hanging pendant lamp gleaming in his dark eyes. This lone source casts a pool of warm, concentrated light onto the scene, creating sharp highlights on his cheekbones and the bridge of his nose while plunging the surrounding pool hall into a soft, blurred darkness. His left hand, firmly splayed on the rich green felt of the table, forms a perfect bridge with the cue stick, the fine traces of pale blue chalk on his fingertips visible against the dark fabric of the tabletop. Steadily and deliberately, he slides the polished maple cue back and forth through his fingers, his gaze unwavering as he moves between the cue ball and the target. His entire posture is one of quiet, focused study, creating an atmosphere of intense intimacy.", + "seed": 18913, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_12/index_5266_seed_18913_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2911355298012495, + "vlm_logprob": -0.2911355298012495, + "vlm_yes_no_margin": 2.175, + "vlm_token_credit": -0.2911355298012495, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8294922113418579 + ], + "opposite_logprob_tokens": [ + -0.5794922113418579 + ], + "logprob_reward": -0.8294922113418579, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8294922113418579, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6859515905380249, + 0.09642638266086578, + 0.04456888139247894, + 0.027950959280133247, + 0.013653062283992767, + 0.01454706210643053, + 0.030429596081376076, + 0.021027138456702232, + 0.022359151393175125, + 0.043086208403110504 + ], + "credit_map": [ + 0.05162809044122696, + 0.040666524320840836, + 0.07411389797925949, + 0.08770762383937836, + 0.10534916073083878, + 0.1337931603193283, + 0.0916505679488182, + 0.10104258358478546, + 0.007028814405202866, + 0.007217157166451216, + 0.012788156978785992, + 0.014499773271381855, + 0.015240594744682312, + 0.0169337447732687, + 0.010841242969036102, + 0.01187689695507288, + 0.004777656868100166, + 0.004497696179896593, + 0.0039309062995016575, + 0.005213791038841009, + 0.008440274745225906, + 0.007828736677765846, + 0.004600520711392164, + 0.005279303062707186, + 0.0036357997450977564, + 0.0035183310974389315, + 0.002692276146262884, + 0.003648563055321574, + 0.004935303702950478, + 0.004927222616970539, + 0.0022425369825214148, + 0.0023509268648922443, + 0.0016648276941850781, + 0.001611563959158957, + 0.0009914718102663755, + 0.0011771096615120769, + 0.0019734897650778294, + 0.0024524927139282227, + 0.001547351130284369, + 0.0022347550839185715, + 0.0014103871071711183, + 0.002163822064176202, + 0.001838532043620944, + 0.0013708758633583784, + 0.0021715937182307243, + 0.002197741298004985, + 0.0017536088125780225, + 0.0016405016649514437, + 0.002409407403320074, + 0.0029910998418927193, + 0.004080634564161301, + 0.005708770826458931, + 0.0038354673888534307, + 0.004410422872751951, + 0.0032317168079316616, + 0.0037620768416672945, + 0.0018683281959965825, + 0.002375561511144042, + 0.0029565237928181887, + 0.0030527927447110415, + 0.002557900734245777, + 0.0029380042105913162, + 0.002132686320692301, + 0.0031453403644263744, + 0.0020852922461926937, + 0.0023961812257766724, + 0.002050141105428338, + 0.0017292246920987964, + 0.0037493209820240736, + 0.00458130519837141, + 0.002382769947871566, + 0.0033849161118268967, + 0.003316489513963461, + 0.003290521912276745, + 0.004115762654691935, + 0.005241048987954855, + 0.005856703966856003, + 0.0066339983604848385, + 0.006262284237891436, + 0.008369401097297668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.8294922113418579 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794747456908226 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.009794747456908226, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794747456908226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7510154843330383, + 0.09297775477170944, + 0.012416099198162556, + 0.011109801940619946, + 0.01120587158948183, + 0.012474847957491875, + 0.03257102146744728, + 0.011578138917684555, + 0.05353965610265732, + 0.011111386120319366 + ], + "credit_map": [ + 0.04930873587727547, + 0.06070245802402496, + 0.07115497440099716, + 0.08485730737447739, + 0.20712709426879883, + 0.19880403578281403, + 0.04425068572163582, + 0.0348101481795311, + 0.007300839759409428, + 0.00826655887067318, + 0.013013132847845554, + 0.013823211193084717, + 0.015388969331979752, + 0.017344096675515175, + 0.00971313752233982, + 0.008127806708216667, + 0.001556833740323782, + 0.0019775941036641598, + 0.001115466351620853, + 0.0011058857198804617, + 0.002376766409724951, + 0.0018884550081565976, + 0.00114467553794384, + 0.0012504213955253363, + 0.0013848838862031698, + 0.0016375775448977947, + 0.0009973437990993261, + 0.001246632426045835, + 0.001868438790552318, + 0.0017073072958737612, + 0.000984605634585023, + 0.0012830119812861085, + 0.0014812690205872059, + 0.0019503802759572864, + 0.0010532692540436983, + 0.0008027998264878988, + 0.0018052523955702782, + 0.0019363999599590898, + 0.0009643682278692722, + 0.0012121321633458138, + 0.0015021961880847812, + 0.0017222997266799212, + 0.0014898378867655993, + 0.0015728796133771539, + 0.0020693009719252586, + 0.0021720416843891144, + 0.0008189024520106614, + 0.0011273896088823676, + 0.003503821324557066, + 0.005451089236885309, + 0.00441450159996748, + 0.00403594272211194, + 0.004594620782881975, + 0.00493021821603179, + 0.0029524497222155333, + 0.0026883771643042564, + 0.0011208312353119254, + 0.0014804109232500196, + 0.0016592593165114522, + 0.0016328557394444942, + 0.0014495158102363348, + 0.0017307689413428307, + 0.0012671890435740352, + 0.0012373077915981412, + 0.004520112182945013, + 0.005659014917910099, + 0.00794358178973198, + 0.006965799257159233, + 0.008978771045804024, + 0.009069987572729588, + 0.005301842000335455, + 0.005100549664348364, + 0.0011873714393004775, + 0.0011877913493663073, + 0.0016382340108975768, + 0.0017226958880200982, + 0.0012439710553735495, + 0.0015053984243422747, + 0.0011649513617157936, + 0.0014609721256420016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.009794747456908226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496501550078392 + ], + "opposite_logprob_tokens": [ + -2.929965019226074 + ], + "logprob_reward": -0.05496501550078392, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496501550078392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4163672924041748, + 0.10506395995616913, + 0.05714152753353119, + 0.04548359289765358, + 0.03378155827522278, + 0.028905559331178665, + 0.055904537439346313, + 0.047138337045907974, + 0.08501269668340683, + 0.12520097196102142 + ], + "credit_map": [ + 0.043390072882175446, + 0.03729362040758133, + 0.04991275817155838, + 0.04873650521039963, + 0.05392847955226898, + 0.07006235420703888, + 0.050397153943777084, + 0.06264633685350418, + 0.009512769058346748, + 0.014769994653761387, + 0.008603890426456928, + 0.013431638479232788, + 0.01371003221720457, + 0.018797479569911957, + 0.011967463418841362, + 0.01427068468183279, + 0.007116984110325575, + 0.00860141683369875, + 0.0039650616236031055, + 0.005589010193943977, + 0.01026390865445137, + 0.009435910731554031, + 0.005282077006995678, + 0.0068871513940393925, + 0.006560340989381075, + 0.007885724306106567, + 0.003238220466300845, + 0.004549909383058548, + 0.007240325678139925, + 0.007997267879545689, + 0.0033063176088035107, + 0.004705479368567467, + 0.003632766893133521, + 0.0062609026208519936, + 0.0026897701900452375, + 0.003215010743588209, + 0.005136233288794756, + 0.00563043775036931, + 0.00319306249730289, + 0.004023372195661068, + 0.002720358781516552, + 0.004982283338904381, + 0.003771699732169509, + 0.003100713947787881, + 0.004185930825769901, + 0.004670930095016956, + 0.0028499288018792868, + 0.002623710548505187, + 0.0038863052614033222, + 0.005511412862688303, + 0.00959703978151083, + 0.012396221049129963, + 0.006287607830017805, + 0.007701315451413393, + 0.004499481059610844, + 0.006025146227329969, + 0.004193101078271866, + 0.005359044764190912, + 0.006634956691414118, + 0.007697205524891615, + 0.006678309757262468, + 0.007356878370046616, + 0.004057551734149456, + 0.005161284003406763, + 0.007019590586423874, + 0.012500777840614319, + 0.012104026973247528, + 0.011120887473225594, + 0.013442282564938068, + 0.01341425534337759, + 0.008504067547619343, + 0.0069068013690412045, + 0.007322010584175587, + 0.007529729045927525, + 0.017614254727959633, + 0.02149975299835205, + 0.01860542595386505, + 0.023488439619541168, + 0.013560832478106022, + 0.015580512583255768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.05496501550078392 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809856250882149 + ], + "opposite_logprob_tokens": [ + -3.2880985736846924 + ], + "logprob_reward": -0.03809856250882149, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809856250882149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4101543426513672, + 0.12643229961395264, + 0.09410898387432098, + 0.044373925775289536, + 0.048145685344934464, + 0.033733971416950226, + 0.06268341094255447, + 0.040078721940517426, + 0.04518755152821541, + 0.09510112553834915 + ], + "credit_map": [ + 0.03069576621055603, + 0.021133728325366974, + 0.031537216156721115, + 0.04364483803510666, + 0.07074899971485138, + 0.08369921147823334, + 0.058571912348270416, + 0.07012267410755157, + 0.015720557421445847, + 0.016499988734722137, + 0.00814297329634428, + 0.010938134044408798, + 0.016961829736828804, + 0.019549453631043434, + 0.01601659692823887, + 0.02260276861488819, + 0.008005276322364807, + 0.010228137485682964, + 0.006560451351106167, + 0.010470982640981674, + 0.013098305091261864, + 0.015975357964634895, + 0.012107336893677711, + 0.01766313426196575, + 0.004804769065231085, + 0.005203246138989925, + 0.002711397595703602, + 0.003707188181579113, + 0.007638600654900074, + 0.007934784516692162, + 0.005016219336539507, + 0.007357715629041195, + 0.0045858765952289104, + 0.004990429617464542, + 0.003443204564973712, + 0.005462155677378178, + 0.005968304816633463, + 0.0075487615540623665, + 0.006905384361743927, + 0.009241566061973572, + 0.003298158524557948, + 0.0038248845376074314, + 0.005419614724814892, + 0.007500799372792244, + 0.0033617897424846888, + 0.004333593882620335, + 0.002998671028763056, + 0.002996461233124137, + 0.005058657377958298, + 0.008184142410755157, + 0.011436857283115387, + 0.014613429084420204, + 0.006676673889160156, + 0.007222281768918037, + 0.004928080830723047, + 0.004563293885439634, + 0.0031733913347125053, + 0.004727653693407774, + 0.004945259541273117, + 0.006128513719886541, + 0.004674731753766537, + 0.005915080662816763, + 0.004504667595028877, + 0.006009423173964024, + 0.004416372161358595, + 0.006343086715787649, + 0.004498662892729044, + 0.0037819917779415846, + 0.008482364006340504, + 0.008720043115317822, + 0.004629182629287243, + 0.004315849393606186, + 0.007549408823251724, + 0.006074925884604454, + 0.011258912272751331, + 0.015114754438400269, + 0.01223839446902275, + 0.017415102571249008, + 0.009740528650581837, + 0.01570909284055233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03809856250882149 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.523327112197876 + ], + "opposite_logprob_tokens": [ + -0.898327112197876 + ], + "logprob_reward": -0.523327112197876, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.523327112197876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5514371991157532, + 0.1320166140794754, + 0.05574371665716171, + 0.028633901849389076, + 0.028855713084340096, + 0.019268618896603584, + 0.03775358945131302, + 0.015140864998102188, + 0.013363084755837917, + 0.11778669059276581 + ], + "credit_map": [ + 0.043242234736680984, + 0.052286531776189804, + 0.050994325429201126, + 0.08054935932159424, + 0.0647052675485611, + 0.08929280936717987, + 0.07384772598743439, + 0.09651892632246017, + 0.01346305850893259, + 0.0127687007188797, + 0.009933900088071823, + 0.014454900287091732, + 0.02020377852022648, + 0.021311815828084946, + 0.016956061124801636, + 0.022924408316612244, + 0.005804166197776794, + 0.007326997816562653, + 0.0038376098964363337, + 0.005144178867340088, + 0.010388422757387161, + 0.009120373986661434, + 0.005752965807914734, + 0.008369001559913158, + 0.003649092512205243, + 0.004262901842594147, + 0.0020799264311790466, + 0.002993975766003132, + 0.004616634454578161, + 0.004808960482478142, + 0.0026785859372466803, + 0.0035438244231045246, + 0.0029915552586317062, + 0.0025921326596289873, + 0.003252970753237605, + 0.003396569984033704, + 0.004191259853541851, + 0.004434541333466768, + 0.003542963182553649, + 0.004453721456229687, + 0.002193630440160632, + 0.002407584572210908, + 0.0018620941555127501, + 0.002581581939011812, + 0.0028761103749275208, + 0.002813593251630664, + 0.0016928657423704863, + 0.002841157605871558, + 0.0038695542607456446, + 0.003564796643331647, + 0.005441711749881506, + 0.00551089970394969, + 0.006242478732019663, + 0.005771768279373646, + 0.003440252272412181, + 0.00391212897375226, + 0.001651255413889885, + 0.0015547366347163916, + 0.001234436989761889, + 0.0013120705261826515, + 0.0025352986995130777, + 0.0024766321294009686, + 0.002018864033743739, + 0.002357570920139551, + 0.0014425339177250862, + 0.0011101487325504422, + 0.00151706428732723, + 0.001522295642644167, + 0.0016653548227623105, + 0.001978729385882616, + 0.0016729586059227586, + 0.0024539988953620195, + 0.012012654915452003, + 0.00932963378727436, + 0.010986234061419964, + 0.014831042848527431, + 0.019595997408032417, + 0.02155415341258049, + 0.014649469405412674, + 0.01482751127332449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.523327112197876 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, tight, eye-level shot, rendered with detailed, grounded realism, focuses on the determined face of an Asian man in his twenties playing pool, leaning over the table to line up a key shot. The camera captures subtle details: the subtle wrinkle of his brow beneath his neatly cropped hair, and the intensely focused reflection of a single, low-hanging pendant lamp gleaming in his dark eyes. This lone source casts a pool of warm, concentrated light onto the scene, creating sharp highlights on his cheekbones and the bridge of his nose while plunging the surrounding pool hall into a soft, blurred darkness. His left hand, firmly splayed on the rich green felt of the table, forms a perfect bridge with the cue stick, the fine traces of pale blue chalk on his fingertips visible against the dark fabric of the tabletop. Steadily and deliberately, he slides the polished maple cue back and forth through his fingers, his gaze unwavering as he moves between the cue ball and the target. His entire posture is one of quiet, focused study, creating an atmosphere of intense intimacy.", + "seed": 18914, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_12/index_5266_seed_18914_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07351830368861556, + "vlm_logprob": -0.07351830368861556, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.07351830368861556, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08351461589336395 + ], + "opposite_logprob_tokens": [ + -2.58351469039917 + ], + "logprob_reward": -0.08351461589336395, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08351461589336395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4385131895542145, + 0.07323827594518661, + 0.06112557277083397, + 0.05767512694001198, + 0.0481097437441349, + 0.0675264224410057, + 0.06538209319114685, + 0.054024286568164825, + 0.12200374901294708, + 0.012401503510773182 + ], + "credit_map": [ + 0.07397176325321198, + 0.0908113643527031, + 0.050101201981306076, + 0.03916410356760025, + 0.10064543783664703, + 0.030381247401237488, + 0.029725512489676476, + 0.023712564259767532, + 0.011666224338114262, + 0.010751559399068356, + 0.007897951640188694, + 0.01194789633154869, + 0.010195707902312279, + 0.0038364308420568705, + 0.006779605522751808, + 0.010162893682718277, + 0.010456081479787827, + 0.009692109189927578, + 0.011863735504448414, + 0.005107671022415161, + 0.010781334713101387, + 0.005313685163855553, + 0.00434834836050868, + 0.0035625998862087727, + 0.009173581376671791, + 0.008802703581750393, + 0.009843782521784306, + 0.006363282445818186, + 0.011652940884232521, + 0.0034276596270501614, + 0.005468904506415129, + 0.002942268969491124, + 0.007428111042827368, + 0.008082186803221703, + 0.006475763861089945, + 0.004855477251112461, + 0.008008906617760658, + 0.00486336275935173, + 0.004752587527036667, + 0.003643343225121498, + 0.010153468698263168, + 0.01272290013730526, + 0.010479697026312351, + 0.008379139006137848, + 0.012231627479195595, + 0.004326292779296637, + 0.00458042835816741, + 0.004652861040085554, + 0.012168489396572113, + 0.006155018694698811, + 0.018546687439084053, + 0.006751437671482563, + 0.007198589853942394, + 0.004141878802329302, + 0.006170062348246574, + 0.004249926190823317, + 0.00926754716783762, + 0.006234301254153252, + 0.013246436603367329, + 0.005622158292680979, + 0.006373141426593065, + 0.0038871364668011665, + 0.004110292531549931, + 0.005283269099891186, + 0.02044115588068962, + 0.014077517203986645, + 0.03229814022779465, + 0.013164572417736053, + 0.01417538896203041, + 0.008419067598879337, + 0.010314226150512695, + 0.009113667532801628, + 0.0021334534976631403, + 0.0014918169472366571, + 0.002611833857372403, + 0.0013898214092478156, + 0.001980548957362771, + 0.0009272359893657267, + 0.0009368208120577037, + 0.0009299717494286597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.08351461589336395 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041139028035104275 + ], + "opposite_logprob_tokens": [ + -5.504113674163818 + ], + "logprob_reward": -0.0041139028035104275, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041139028035104275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34393930435180664, + 0.06963963806629181, + 0.1179441511631012, + 0.04060179740190506, + 0.05785627290606499, + 0.04995475336909294, + 0.08824694901704788, + 0.07297280430793762, + 0.13254503905773163, + 0.026299381628632545 + ], + "credit_map": [ + 0.05855794996023178, + 0.06671984493732452, + 0.0399724580347538, + 0.0345490388572216, + 0.0775575339794159, + 0.02331572398543358, + 0.024420734494924545, + 0.018845954909920692, + 0.010251070372760296, + 0.013286136090755463, + 0.008318128064274788, + 0.006772923748940229, + 0.012466631829738617, + 0.007357045542448759, + 0.005808493122458458, + 0.005379204638302326, + 0.021129747852683067, + 0.013084867969155312, + 0.030689982697367668, + 0.013172052800655365, + 0.012846527621150017, + 0.006311072036623955, + 0.014134807512164116, + 0.006575082894414663, + 0.007186216302216053, + 0.006435273680835962, + 0.005863754544407129, + 0.005839588586241007, + 0.007160527631640434, + 0.0022376792039722204, + 0.0027904589660465717, + 0.0030882901046425104, + 0.008948715403676033, + 0.009244491346180439, + 0.006458553019911051, + 0.006713303737342358, + 0.008895257487893105, + 0.010225369594991207, + 0.0037336149252951145, + 0.0036369538865983486, + 0.007601880934089422, + 0.005500405095517635, + 0.006979731377214193, + 0.0091525474563241, + 0.006194266024976969, + 0.004469084553420544, + 0.004462589975446463, + 0.005594244692474604, + 0.015259623527526855, + 0.012190982699394226, + 0.01344109047204256, + 0.012027383781969547, + 0.01329083926975727, + 0.006486092694103718, + 0.006671146955341101, + 0.008879778906702995, + 0.01194548886269331, + 0.008230846375226974, + 0.01669217087328434, + 0.008079159073531628, + 0.008870630525052547, + 0.005443984642624855, + 0.007378975860774517, + 0.006331542041152716, + 0.02244045026600361, + 0.014861918054521084, + 0.036054618656635284, + 0.013909549452364445, + 0.015008820220828056, + 0.008881130255758762, + 0.011208541691303253, + 0.010179983451962471, + 0.004769146908074617, + 0.0035733231343328953, + 0.005957703571766615, + 0.002575973281636834, + 0.003554754424840212, + 0.001782208331860602, + 0.002147227292880416, + 0.001939042005687952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0041139028035104275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979751117527485 + ], + "opposite_logprob_tokens": [ + -3.5297975540161133 + ], + "logprob_reward": -0.02979751117527485, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979751117527485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24693924188613892, + 0.1111622005701065, + 0.14052090048789978, + 0.08072501420974731, + 0.2086174637079239, + 0.04904014244675636, + 0.040578603744506836, + 0.04858289286494255, + 0.05525774136185646, + 0.018575815483927727 + ], + "credit_map": [ + 0.04210219532251358, + 0.05272423475980759, + 0.025049008429050446, + 0.025993043556809425, + 0.05215334892272949, + 0.011552531272172928, + 0.017155224457383156, + 0.0202096588909626, + 0.015780756250023842, + 0.017696410417556763, + 0.014795233495533466, + 0.015393991023302078, + 0.017572611570358276, + 0.01337484922260046, + 0.007387310266494751, + 0.009161041118204594, + 0.024346260353922844, + 0.021248377859592438, + 0.024700630456209183, + 0.016968591138720512, + 0.021901562809944153, + 0.011630061082541943, + 0.009651219472289085, + 0.010074194520711899, + 0.013985667377710342, + 0.013770082034170628, + 0.01337946392595768, + 0.009634504094719887, + 0.01516036968678236, + 0.005340792238712311, + 0.005222167819738388, + 0.004231971688568592, + 0.030815303325653076, + 0.036393776535987854, + 0.02093588002026081, + 0.015988174825906754, + 0.03894035145640373, + 0.036721404641866684, + 0.01280796155333519, + 0.016014626249670982, + 0.008112093433737755, + 0.007308688946068287, + 0.009083718992769718, + 0.006865744944661856, + 0.006130036897957325, + 0.004315334837883711, + 0.00388364982791245, + 0.0033408752642571926, + 0.007261454593390226, + 0.005333069711923599, + 0.008577864617109299, + 0.003983666189014912, + 0.005515755154192448, + 0.002880035201087594, + 0.0026661918964236975, + 0.004360564053058624, + 0.008342885412275791, + 0.006427144166082144, + 0.011631696484982967, + 0.004470045678317547, + 0.007261388003826141, + 0.0032697266433387995, + 0.0034907343797385693, + 0.0036892713978886604, + 0.009245146997272968, + 0.006248772144317627, + 0.01562812365591526, + 0.00553768128156662, + 0.007054157089442015, + 0.0036717127077281475, + 0.0045254165306687355, + 0.0033467309549450874, + 0.003007001942023635, + 0.0027290182188153267, + 0.002141817705705762, + 0.0022367655765265226, + 0.004180580377578735, + 0.001058485358953476, + 0.0016794322291389108, + 0.0015427138423547149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.02979751117527485 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864047095179558 + ], + "opposite_logprob_tokens": [ + -3.048640489578247 + ], + "logprob_reward": -0.04864047095179558, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864047095179558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2656788229942322, + 0.12527385354042053, + 0.07675362378358841, + 0.21734268963336945, + 0.07923278212547302, + 0.04085483029484749, + 0.04110826924443245, + 0.04525931552052498, + 0.07738654315471649, + 0.031109245494008064 + ], + "credit_map": [ + 0.04414074495434761, + 0.045614924281835556, + 0.03324223309755325, + 0.022699180990457535, + 0.0498894602060318, + 0.02820516750216484, + 0.0163775235414505, + 0.025509605184197426, + 0.020568598061800003, + 0.020759621635079384, + 0.013106780126690865, + 0.010004239156842232, + 0.02200976572930813, + 0.018636472523212433, + 0.008133037947118282, + 0.012055344879627228, + 0.011330394074320793, + 0.009867367334663868, + 0.01750222034752369, + 0.009399070404469967, + 0.009597638621926308, + 0.00436545442789793, + 0.00840273592621088, + 0.006288744509220123, + 0.03580030798912048, + 0.028794724494218826, + 0.048928093165159225, + 0.02478836476802826, + 0.03432142361998558, + 0.012457777746021748, + 0.020456450060009956, + 0.01179554220288992, + 0.012174291536211967, + 0.011231777258217335, + 0.00860954262316227, + 0.008392160758376122, + 0.012839904054999352, + 0.01376780029386282, + 0.004847642965614796, + 0.007369667757302523, + 0.006518763490021229, + 0.007261347956955433, + 0.005047062877565622, + 0.0035049004945904016, + 0.007222883868962526, + 0.003202648600563407, + 0.0028450279496610165, + 0.005252195987850428, + 0.007156672887504101, + 0.006310217082500458, + 0.007416090928018093, + 0.0037446534261107445, + 0.0056935641914606094, + 0.0031302825082093477, + 0.0031025060452520847, + 0.004554280545562506, + 0.007637140341103077, + 0.007117862813174725, + 0.006571928504854441, + 0.0052336761727929115, + 0.007738469168543816, + 0.004553052596747875, + 0.003253546077758074, + 0.003153638681396842, + 0.012729562819004059, + 0.00993930734694004, + 0.014773134142160416, + 0.008508960716426373, + 0.01300173532217741, + 0.006109666079282761, + 0.005444391164928675, + 0.006879784632474184, + 0.0054939184337854385, + 0.00543863233178854, + 0.005709745921194553, + 0.002710177795961499, + 0.005133106838911772, + 0.002427350962534547, + 0.0019567031413316727, + 0.0022396105341613293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.04864047095179558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.201525017619133 + ], + "opposite_logprob_tokens": [ + -1.7015249729156494 + ], + "logprob_reward": -0.201525017619133, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.201525017619133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39947453141212463, + 0.09938225895166397, + 0.09685084223747253, + 0.09295583516359329, + 0.04528505727648735, + 0.08584282547235489, + 0.05275392159819603, + 0.028214814141392708, + 0.06869947165250778, + 0.03054036758840084 + ], + "credit_map": [ + 0.06680668145418167, + 0.08508580178022385, + 0.044633474200963974, + 0.035202961415052414, + 0.09149065613746643, + 0.026022333651781082, + 0.026019880548119545, + 0.02421276830136776, + 0.01504692155867815, + 0.020649360492825508, + 0.012483496218919754, + 0.012147614732384682, + 0.02019529975950718, + 0.006347788963466883, + 0.006301598157733679, + 0.0062101841904222965, + 0.013977166265249252, + 0.01083388365805149, + 0.021928830072283745, + 0.0124860480427742, + 0.013983341865241528, + 0.006045397371053696, + 0.010628161020576954, + 0.0069680167362093925, + 0.01667078398168087, + 0.010952357202768326, + 0.022699197754263878, + 0.012319289147853851, + 0.011483576148748398, + 0.005946777295321226, + 0.007662431336939335, + 0.0052214330062270164, + 0.007469584234058857, + 0.0071558877825737, + 0.007473147474229336, + 0.005004527512937784, + 0.005881706718355417, + 0.0034883737098425627, + 0.004241087008267641, + 0.004570748191326857, + 0.012738668359816074, + 0.012941312976181507, + 0.016873568296432495, + 0.011067572981119156, + 0.012380985543131828, + 0.0049785771407186985, + 0.006650210823863745, + 0.008211935870349407, + 0.00831473246216774, + 0.007703364826738834, + 0.0052200546488165855, + 0.0073348828591406345, + 0.010566124692559242, + 0.004458037205040455, + 0.005379937123507261, + 0.0037767926696687937, + 0.004743651952594519, + 0.004137455020099878, + 0.004833483602851629, + 0.002655448392033577, + 0.005579677876085043, + 0.0017318538157269359, + 0.002026171423494816, + 0.0025070742703974247, + 0.011222541332244873, + 0.009336617775261402, + 0.009931661188602448, + 0.008843946270644665, + 0.013582824729382992, + 0.00499785877764225, + 0.006642989814281464, + 0.004141037352383137, + 0.005316736176609993, + 0.005219246260821819, + 0.0031258186791092157, + 0.003833552822470665, + 0.0072444346733391285, + 0.001744092209264636, + 0.002198915695771575, + 0.0018575743306428194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.201525017619133 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, tight, eye-level shot, rendered with detailed, grounded realism, focuses on the determined face of an Asian man in his twenties playing pool, leaning over the table to line up a key shot. The camera captures subtle details: the subtle wrinkle of his brow beneath his neatly cropped hair, and the intensely focused reflection of a single, low-hanging pendant lamp gleaming in his dark eyes. This lone source casts a pool of warm, concentrated light onto the scene, creating sharp highlights on his cheekbones and the bridge of his nose while plunging the surrounding pool hall into a soft, blurred darkness. His left hand, firmly splayed on the rich green felt of the table, forms a perfect bridge with the cue stick, the fine traces of pale blue chalk on his fingertips visible against the dark fabric of the tabletop. Steadily and deliberately, he slides the polished maple cue back and forth through his fingers, his gaze unwavering as he moves between the cue ball and the target. His entire posture is one of quiet, focused study, creating an atmosphere of intense intimacy.", + "seed": 18915, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_12/index_5266_seed_18915_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08505753381177783, + "vlm_logprob": -0.08505753381177783, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.08505753381177783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07403673231601715 + ], + "opposite_logprob_tokens": [ + -2.6990368366241455 + ], + "logprob_reward": -0.07403673231601715, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07403673231601715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1775769144296646, + 0.4321640133857727, + 0.02017761580646038, + 0.0147673599421978, + 0.008679384365677834, + 0.021252430975437164, + 0.02461812272667885, + 0.01549056638032198, + 0.005630395375192165, + 0.27964311838150024 + ], + "credit_map": [ + 0.011708424426615238, + 0.040945865213871, + 0.029559332877397537, + 0.015669923275709152, + 0.027124613523483276, + 0.03583814576268196, + 0.008722526952624321, + 0.008008060976862907, + 0.014824566431343555, + 0.02933800034224987, + 0.014207501895725727, + 0.005916398484259844, + 0.17762237787246704, + 0.17416875064373016, + 0.004105919040739536, + 0.011980527080595493, + 0.002318001352250576, + 0.0038779503665864468, + 0.002958759432658553, + 0.0025714533403515816, + 0.00315987435169518, + 0.002634153701364994, + 0.0013035539304837584, + 0.0013538679340854287, + 0.0016039058100432158, + 0.0035290424712002277, + 0.0018604157958179712, + 0.0012990471441298723, + 0.0024625493679195642, + 0.0020118975080549717, + 0.0008909470052458346, + 0.0011095539666712284, + 0.0008918591775000095, + 0.0013708443148061633, + 0.0014608820201829076, + 0.001517611206509173, + 0.0009963406482711434, + 0.001055702450685203, + 0.000701064127497375, + 0.000685079547110945, + 0.0022898539900779724, + 0.002445365535095334, + 0.003164764028042555, + 0.003178635612130165, + 0.003845581319183111, + 0.0027694872114807367, + 0.0012635577004402876, + 0.002295185811817646, + 0.0016087591648101807, + 0.002829113509505987, + 0.0024825618602335453, + 0.0024566284846514463, + 0.005761823151260614, + 0.006832966115325689, + 0.00116734451148659, + 0.001478925347328186, + 0.0012319452362135053, + 0.0023671556264162064, + 0.0028746221214532852, + 0.002268236130475998, + 0.001743415603414178, + 0.0023629122879356146, + 0.0012305862037464976, + 0.001411694218404591, + 0.0005575459799729288, + 0.0011036007199436426, + 0.000552790064830333, + 0.0005205406923778355, + 0.001231963629834354, + 0.0009052401874214411, + 0.0003493993717711419, + 0.0004093149909749627, + 0.012359652668237686, + 0.016617514193058014, + 0.030541619285941124, + 0.0291589368134737, + 0.08160872012376785, + 0.08422791957855225, + 0.01013548206537962, + 0.01499329786747694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.07403673231601715 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036309524439275265 + ], + "opposite_logprob_tokens": [ + -5.628631114959717 + ], + "logprob_reward": -0.0036309524439275265, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036309524439275265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15128307044506073, + 0.16571958363056183, + 0.029561804607510567, + 0.021844595670700073, + 0.01777968741953373, + 0.109842948615551, + 0.011647756211459637, + 0.007913313806056976, + 0.015446514822542667, + 0.4689607620239258 + ], + "credit_map": [ + 0.009063851088285446, + 0.020343733951449394, + 0.031939197331666946, + 0.01333623006939888, + 0.018461376428604126, + 0.03357759490609169, + 0.01308689545840025, + 0.01147419773042202, + 0.005771424621343613, + 0.012677990831434727, + 0.004932854790240526, + 0.004644615110009909, + 0.06829658895730972, + 0.06272551417350769, + 0.0020822423975914717, + 0.004588358104228973, + 0.0025489977560937405, + 0.007160605862736702, + 0.0026203745510429144, + 0.002249440411105752, + 0.005198467522859573, + 0.005121663212776184, + 0.0025235386565327644, + 0.002138715935871005, + 0.0020178919658064842, + 0.00574229471385479, + 0.0022971222642809153, + 0.0018207000102847815, + 0.003840936813503504, + 0.004166614264249802, + 0.0009399121045134962, + 0.00101912347599864, + 0.002557365456596017, + 0.003070814535021782, + 0.0020076546352356672, + 0.001617806381545961, + 0.0030295897740870714, + 0.0031782607547938824, + 0.0009640007629059255, + 0.0013541948283091187, + 0.009487230330705643, + 0.008001874200999737, + 0.023013409227132797, + 0.023194415494799614, + 0.016459506005048752, + 0.02167557179927826, + 0.004437100607901812, + 0.003573847934603691, + 0.0007129819132387638, + 0.002523651346564293, + 0.0008792506996542215, + 0.0012145631480962038, + 0.0022847186774015427, + 0.002856239676475525, + 0.0006212054868228734, + 0.0005551458452828228, + 0.0008124650921672583, + 0.0012935919221490622, + 0.0015608182875439525, + 0.001206613495014608, + 0.0009589880937710404, + 0.0009690177976153791, + 0.0005799189675599337, + 0.0005319009069353342, + 0.0007016913732513785, + 0.001391405239701271, + 0.001123924506828189, + 0.0009962881449609995, + 0.004904357250779867, + 0.005087296012789011, + 0.0006208542035892606, + 0.0006206976249814034, + 0.017278268933296204, + 0.029314175248146057, + 0.035480670630931854, + 0.03212017938494682, + 0.1735791563987732, + 0.16344821453094482, + 0.008272428996860981, + 0.009467629715800285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0036309524439275265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288656026124954 + ], + "opposite_logprob_tokens": [ + -3.7732887268066406 + ], + "logprob_reward": -0.023288656026124954, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288656026124954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16891968250274658, + 0.3752036690711975, + 0.03042515739798546, + 0.03612387180328369, + 0.06281626969575882, + 0.1353687196969986, + 0.00910951942205429, + 0.013336548581719398, + 0.01129588671028614, + 0.15740066766738892 + ], + "credit_map": [ + 0.011547987349331379, + 0.018224850296974182, + 0.03161942958831787, + 0.015295830555260181, + 0.02010609582066536, + 0.03720191866159439, + 0.019019341096282005, + 0.01590423472225666, + 0.013126675970852375, + 0.025701358914375305, + 0.012515311129391193, + 0.0049764234572649, + 0.15461893379688263, + 0.1502406895160675, + 0.0035475294571369886, + 0.010476714931428432, + 0.0037733162753283978, + 0.006989059504121542, + 0.003603044431656599, + 0.0030868512112647295, + 0.0044813998974859715, + 0.004261874593794346, + 0.0019962117075920105, + 0.0022333983797580004, + 0.0055893538519740105, + 0.007246237713843584, + 0.004475454334169626, + 0.0045923045836389065, + 0.005691857542842627, + 0.004800172057002783, + 0.0015962693141773343, + 0.002132221357896924, + 0.008586146868765354, + 0.010222293436527252, + 0.007017309311777353, + 0.005532948300242424, + 0.011754915118217468, + 0.012603416107594967, + 0.0027640576008707285, + 0.004335182253271341, + 0.012329758144915104, + 0.010833586566150188, + 0.028434645384550095, + 0.028286857530474663, + 0.018308326601982117, + 0.02521008439362049, + 0.00587734580039978, + 0.0060881199315190315, + 0.0008912617922760546, + 0.001551936729811132, + 0.0013358896831050515, + 0.0016095417086035013, + 0.0013858187012374401, + 0.0012420832645148039, + 0.0005771462456323206, + 0.0005158410058356822, + 0.0014314817963168025, + 0.002802437637001276, + 0.0017845515394583344, + 0.0017043983098119497, + 0.00209094793535769, + 0.0016598833026364446, + 0.0009582725469954312, + 0.0009045753395184875, + 0.0006408295012079179, + 0.0009601440979167819, + 0.0008022726397030056, + 0.0008046816801652312, + 0.0033644528593868017, + 0.0032706987112760544, + 0.0006573923747055233, + 0.0007954146713018417, + 0.007899043150246143, + 0.012988032773137093, + 0.004477821756154299, + 0.004504960495978594, + 0.06096409261226654, + 0.05254923924803734, + 0.005822307895869017, + 0.008195161819458008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.023288656026124954 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304946959018707 + ], + "opposite_logprob_tokens": [ + -3.1680495738983154 + ], + "logprob_reward": -0.04304946959018707, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304946959018707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014978948049247265, + 0.7228899002075195, + 0.033803872764110565, + 0.015543976798653603, + 0.011380279436707497, + 0.03277739882469177, + 0.0030172078404575586, + 0.0016890713013708591, + 0.007024495862424374, + 0.15689486265182495 + ], + "credit_map": [ + 0.0007423929637297988, + 0.0016683327266946435, + 0.003639416303485632, + 0.001868356135673821, + 0.0015684721292927861, + 0.0036131758242845535, + 0.0008001916576176882, + 0.0010786086786538363, + 0.0246132742613554, + 0.050868939608335495, + 0.019264711067080498, + 0.008950349874794483, + 0.30194422602653503, + 0.2907082736492157, + 0.006671377923339605, + 0.019868766888976097, + 0.00248711952008307, + 0.004729049280285835, + 0.0039021861739456654, + 0.003989303018897772, + 0.006158828269690275, + 0.0069613223895430565, + 0.0030557229183614254, + 0.002520341891795397, + 0.0011936321388930082, + 0.0030525624752044678, + 0.0019548845011740923, + 0.001579135307110846, + 0.0027564745396375656, + 0.0026974815409630537, + 0.0012321746908128262, + 0.0010776297422125936, + 0.0010035410523414612, + 0.0017565494636073709, + 0.0011030216701328754, + 0.0009334243950434029, + 0.002430424327030778, + 0.002485761884599924, + 0.0006314333877526224, + 0.00103612313978374, + 0.0026776485610753298, + 0.0023077945224940777, + 0.007056293077766895, + 0.0068730697967112064, + 0.004984468687325716, + 0.006590916775166988, + 0.0012920321896672249, + 0.000995173119008541, + 0.0002608744543977082, + 0.00038363190833479166, + 0.0003857339615933597, + 0.0005186854396015406, + 0.0005611711530946195, + 0.0005858248332515359, + 0.00015864749730098993, + 0.0001626386947464198, + 0.00013767079508397728, + 0.0002565197937656194, + 0.00023244581825565547, + 0.0002646290813572705, + 0.0002253918064525351, + 0.0002466676523908973, + 0.00015738928050268441, + 0.00016835711721796542, + 0.0005835672491230071, + 0.0007966588600538671, + 0.0008001321693882346, + 0.0007644800934940577, + 0.0016343328170478344, + 0.0016694311052560806, + 0.00038529143785126507, + 0.00039060221752151847, + 0.005933435633778572, + 0.009643331170082092, + 0.01533091627061367, + 0.014010699465870857, + 0.05416135489940643, + 0.05148228630423546, + 0.0031095880549401045, + 0.003223237581551075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.04304946959018707 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28128185868263245 + ], + "opposite_logprob_tokens": [ + -1.40628182888031 + ], + "logprob_reward": -0.28128185868263245, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28128185868263245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08918184787034988, + 0.469912052154541, + 0.026307297870516777, + 0.021159471943974495, + 0.015710782259702682, + 0.23061713576316833, + 0.02830454520881176, + 0.011119269765913486, + 0.007700242102146149, + 0.09998738020658493 + ], + "credit_map": [ + 0.006368831731379032, + 0.009155000559985638, + 0.01416748110204935, + 0.007944337092339993, + 0.013623068109154701, + 0.021616337820887566, + 0.008492086082696915, + 0.007814701646566391, + 0.016084259375929832, + 0.03379470854997635, + 0.013139840215444565, + 0.006309990305453539, + 0.19468897581100464, + 0.18736985325813293, + 0.00554439565166831, + 0.012980040162801743, + 0.003213565330952406, + 0.004754394292831421, + 0.004053474869579077, + 0.003314661094918847, + 0.0030774238985031843, + 0.003771797753870487, + 0.002190589439123869, + 0.0019313916563987732, + 0.0018480246653780341, + 0.005337570793926716, + 0.0020890492014586926, + 0.0023371523711830378, + 0.003485205350443721, + 0.0030788746662437916, + 0.0015128368977457285, + 0.0014707592781633139, + 0.0014338616747409105, + 0.0029217570554465055, + 0.0015799106331542134, + 0.0010772545356303453, + 0.00346720521338284, + 0.003590106265619397, + 0.0006005793693475425, + 0.0010401064064353704, + 0.018716659396886826, + 0.015993699431419373, + 0.048491403460502625, + 0.04783271625638008, + 0.03578381612896919, + 0.047438185662031174, + 0.008980258367955685, + 0.007380393799394369, + 0.0022410289384424686, + 0.004565001931041479, + 0.004047587979584932, + 0.003592872293666005, + 0.004985329229384661, + 0.005342935677617788, + 0.0015639241319149733, + 0.001965863164514303, + 0.0007486551767215133, + 0.0015199503395706415, + 0.0014685442438349128, + 0.0013560978695750237, + 0.0020106318406760693, + 0.001796923577785492, + 0.0010952252196148038, + 0.0011232407996430993, + 0.0008615299011580646, + 0.0014833103632554412, + 0.0008248030208051205, + 0.001048112171702087, + 0.0011464967392385006, + 0.0011169321369379759, + 0.0006398924742825329, + 0.0005791656440123916, + 0.00875828880816698, + 0.010846125893294811, + 0.01792961359024048, + 0.018748173490166664, + 0.010900798253715038, + 0.0110565684735775, + 0.008798652328550816, + 0.012949160300195217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.28128185868263245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, tight, eye-level shot, rendered with detailed, grounded realism, focuses on the determined face of an Asian man in his twenties playing pool, leaning over the table to line up a key shot. The camera captures subtle details: the subtle wrinkle of his brow beneath his neatly cropped hair, and the intensely focused reflection of a single, low-hanging pendant lamp gleaming in his dark eyes. This lone source casts a pool of warm, concentrated light onto the scene, creating sharp highlights on his cheekbones and the bridge of his nose while plunging the surrounding pool hall into a soft, blurred darkness. His left hand, firmly splayed on the rich green felt of the table, forms a perfect bridge with the cue stick, the fine traces of pale blue chalk on his fingertips visible against the dark fabric of the tabletop. Steadily and deliberately, he slides the polished maple cue back and forth through his fingers, his gaze unwavering as he moves between the cue ball and the target. His entire posture is one of quiet, focused study, creating an atmosphere of intense intimacy.", + "seed": 18916, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_12/index_5266_seed_18916_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04621708067134023, + "vlm_logprob": -0.04621708067134023, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.04621708067134023, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07517454773187637 + ], + "opposite_logprob_tokens": [ + -2.700174570083618 + ], + "logprob_reward": -0.07517454773187637, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07517454773187637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1771332174539566, + 0.050836771726608276, + 0.025684626772999763, + 0.055052973330020905, + 0.04590415582060814, + 0.016214242205023766, + 0.029765937477350235, + 0.05095316097140312, + 0.06471823155879974, + 0.48373669385910034 + ], + "credit_map": [ + 0.021411675959825516, + 0.040053922683000565, + 0.01926703192293644, + 0.02088705077767372, + 0.01912367343902588, + 0.028054000809788704, + 0.014581568539142609, + 0.013754310086369514, + 0.0073979995213449, + 0.01200371328741312, + 0.005103212781250477, + 0.003937745466828346, + 0.006166236940771341, + 0.008741023950278759, + 0.003154302714392543, + 0.0043325405567884445, + 0.0038650836795568466, + 0.005669901147484779, + 0.0026380226481705904, + 0.0023180947173386812, + 0.002887918148189783, + 0.004416085779666901, + 0.0018528330838307738, + 0.0020366893149912357, + 0.010673755779862404, + 0.013309129513800144, + 0.004982804413884878, + 0.004564447328448296, + 0.005547941196709871, + 0.010544225573539734, + 0.002451015170663595, + 0.002979656681418419, + 0.005863700993359089, + 0.011188039556145668, + 0.00551936449483037, + 0.0035902659874409437, + 0.004970801994204521, + 0.008608237840235233, + 0.003376908600330353, + 0.0027868354227393866, + 0.00250010727904737, + 0.003345571458339691, + 0.0016940817004069686, + 0.0010697308462113142, + 0.002685623476281762, + 0.0027384059503674507, + 0.0010926557006314397, + 0.0010880674235522747, + 0.0037926461081951857, + 0.008312920108437538, + 0.002010797383263707, + 0.0024420437403023243, + 0.003254940500482917, + 0.0052942982874810696, + 0.001969407545402646, + 0.0026888849679380655, + 0.0089946910738945, + 0.01337661687284708, + 0.004075012635439634, + 0.0041490588337183, + 0.004374654497951269, + 0.009551911614835262, + 0.0036719534546136856, + 0.0027592689730226994, + 0.009141966700553894, + 0.018964650109410286, + 0.0052605196833610535, + 0.004737964365631342, + 0.006606729235500097, + 0.013050207868218422, + 0.003996498417109251, + 0.0029596968088299036, + 0.06625834107398987, + 0.15402425825595856, + 0.03589368611574173, + 0.02907777577638626, + 0.045826248824596405, + 0.10194822400808334, + 0.029165102168917656, + 0.02154308743774891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.07517454773187637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110934678465128 + ], + "opposite_logprob_tokens": [ + -5.504110813140869 + ], + "logprob_reward": -0.004110934678465128, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110934678465128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2984255850315094, + 0.19814234972000122, + 0.0728946104645729, + 0.07508350908756256, + 0.05341409146785736, + 0.03129812702536583, + 0.07239699363708496, + 0.035104960203170776, + 0.04630443453788757, + 0.1169353723526001 + ], + "credit_map": [ + 0.04021143913269043, + 0.04964270442724228, + 0.039318181574344635, + 0.03647051006555557, + 0.028147853910923004, + 0.042098965495824814, + 0.03956757113337517, + 0.022968340665102005, + 0.031000087037682533, + 0.04821546748280525, + 0.015299554914236069, + 0.017476236447691917, + 0.025245964527130127, + 0.036446746438741684, + 0.013360475189983845, + 0.011097807437181473, + 0.013978835195302963, + 0.016886170953512192, + 0.006558135151863098, + 0.005411893595010042, + 0.007203476503491402, + 0.014087190851569176, + 0.003984163980931044, + 0.004784745629876852, + 0.00782360415905714, + 0.011378756724298, + 0.009296728298068047, + 0.006967016030102968, + 0.010810033418238163, + 0.011790432967245579, + 0.008577730506658554, + 0.008439202792942524, + 0.006429758388549089, + 0.012445429340004921, + 0.00627488549798727, + 0.0040898509323596954, + 0.006889889016747475, + 0.009081751108169556, + 0.004216328263282776, + 0.003986199852079153, + 0.003938669804483652, + 0.007619814481586218, + 0.0041767205111682415, + 0.002384466351941228, + 0.0034903008490800858, + 0.005176665727049112, + 0.0024858100805431604, + 0.002025680849328637, + 0.009133173152804375, + 0.017040789127349854, + 0.006948487833142281, + 0.0064900717698037624, + 0.009119942784309387, + 0.011900847777724266, + 0.00576064782217145, + 0.006003034301102161, + 0.0038920671213418245, + 0.009461996145546436, + 0.0037185028195381165, + 0.0023198984563350677, + 0.004721314180642366, + 0.0064445557072758675, + 0.0020133634097874165, + 0.0025332632940262556, + 0.006404414307326078, + 0.013656921684741974, + 0.003502199426293373, + 0.0038509410806000233, + 0.003747261129319668, + 0.008664627559483051, + 0.0035813951399177313, + 0.0028966758400201797, + 0.016196202486753464, + 0.028640074655413628, + 0.013023610226809978, + 0.007911209017038345, + 0.01130188349634409, + 0.019809717312455177, + 0.008467298932373524, + 0.011585370637476444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.004110934678465128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430544950067997 + ], + "opposite_logprob_tokens": [ + -3.1680545806884766 + ], + "logprob_reward": -0.0430544950067997, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430544950067997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2420811802148819, + 0.12132727354764938, + 0.03615235537290573, + 0.05631311237812042, + 0.1663997918367386, + 0.0498138926923275, + 0.06045529618859291, + 0.06569904088973999, + 0.05748094618320465, + 0.14427706599235535 + ], + "credit_map": [ + 0.025396686047315598, + 0.030965302139520645, + 0.04252229258418083, + 0.04178537428379059, + 0.023808283731341362, + 0.033006686717271805, + 0.02491719275712967, + 0.019679371267557144, + 0.01938171498477459, + 0.0275319404900074, + 0.009118380025029182, + 0.011745636351406574, + 0.014698045328259468, + 0.02230590395629406, + 0.008118224330246449, + 0.008427430875599384, + 0.005058367270976305, + 0.007281459867954254, + 0.004564570728689432, + 0.003313956782221794, + 0.0035904853139072657, + 0.0058125765062868595, + 0.0033075460232794285, + 0.003223390318453312, + 0.007855994626879692, + 0.011688662692904472, + 0.008431126363575459, + 0.0042958930134773254, + 0.005771379452198744, + 0.009493215009570122, + 0.002926543354988098, + 0.0058502983301877975, + 0.0224737711250782, + 0.039569396525621414, + 0.016891399398446083, + 0.012272103689610958, + 0.019351063296198845, + 0.029850035905838013, + 0.013584833592176437, + 0.01240717712789774, + 0.008525860495865345, + 0.012234500609338284, + 0.005406007170677185, + 0.0037483260966837406, + 0.005297428462654352, + 0.008602243848145008, + 0.0031700313556939363, + 0.0028294967487454414, + 0.009403691627085209, + 0.014567489735782146, + 0.005878064781427383, + 0.005793931894004345, + 0.004987877793610096, + 0.01020263321697712, + 0.005270917434245348, + 0.004350688774138689, + 0.009364833123981953, + 0.020585665479302406, + 0.005607043392956257, + 0.004194118548184633, + 0.006255740765482187, + 0.01135045476257801, + 0.0039986297488212585, + 0.004342557396739721, + 0.0070841433480381966, + 0.014550762251019478, + 0.004941478371620178, + 0.005385836120694876, + 0.007171133067458868, + 0.01025763526558876, + 0.004969421308487654, + 0.0031205371487885714, + 0.02273014560341835, + 0.03921158239245415, + 0.014787589199841022, + 0.007281808648258448, + 0.016222218051552773, + 0.02682754024863243, + 0.007946409285068512, + 0.009269770234823227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.0430544950067997 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029790684580802917 + ], + "opposite_logprob_tokens": [ + -3.5297906398773193 + ], + "logprob_reward": -0.029790684580802917, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029790684580802917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21915484964847565, + 0.09221360832452774, + 0.15875622630119324, + 0.0882311761379242, + 0.05050928145647049, + 0.05457070469856262, + 0.047886479645967484, + 0.045355647802352905, + 0.037476796656847, + 0.20584526658058167 + ], + "credit_map": [ + 0.025516919791698456, + 0.05109404772520065, + 0.019936980679631233, + 0.021668849512934685, + 0.025212056934833527, + 0.03604935109615326, + 0.026098649948835373, + 0.013578001409769058, + 0.014892986044287682, + 0.017380278557538986, + 0.009677354246377945, + 0.00766087556257844, + 0.012480086646974087, + 0.014138270169496536, + 0.007256140932440758, + 0.00872761756181717, + 0.019416509196162224, + 0.03327144682407379, + 0.01536332257091999, + 0.017780721187591553, + 0.015963461250066757, + 0.02530362829566002, + 0.017229778692126274, + 0.014427357353270054, + 0.015046615153551102, + 0.01894949935376644, + 0.009238018654286861, + 0.008662085980176926, + 0.008979200385510921, + 0.015840210020542145, + 0.006443686317652464, + 0.00507186958566308, + 0.0083929980173707, + 0.008635344915091991, + 0.007798158563673496, + 0.004214425105601549, + 0.0056076799519360065, + 0.008970172144472599, + 0.0034413179382681847, + 0.0034491869155317545, + 0.009452315047383308, + 0.011561999097466469, + 0.005293893627822399, + 0.0045774164609611034, + 0.006442518904805183, + 0.009125787764787674, + 0.004410260356962681, + 0.0037065118085592985, + 0.007012973539531231, + 0.006477898918092251, + 0.005961512681096792, + 0.0051016369834542274, + 0.00579438079148531, + 0.007106298580765724, + 0.0052745528519153595, + 0.0051572285592556, + 0.005772328469902277, + 0.010960076004266739, + 0.0050312732346355915, + 0.004264766816049814, + 0.0053326645866036415, + 0.007114693056792021, + 0.002791984938085079, + 0.004087862093001604, + 0.0051270523108541965, + 0.006534558720886707, + 0.004725680686533451, + 0.003481683088466525, + 0.004323354456573725, + 0.0059395432472229, + 0.003664526389911771, + 0.0036804007831960917, + 0.029918024316430092, + 0.05637599155306816, + 0.016683116555213928, + 0.01014115009456873, + 0.03226235508918762, + 0.03656848520040512, + 0.01082513015717268, + 0.013070996850728989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.029790684580802917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man in his twenties with neatly cropped hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pool hall with a rich green felt table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single low-hanging pendant lamp creating sharp highlights on the man's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man's left hand visible with pale blue chalk traces on his fingertips while holding a cue stick?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a tight shot focusing on the man's face and upper body?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895474135875702 + ], + "opposite_logprob_tokens": [ + -2.5789546966552734 + ], + "logprob_reward": -0.07895474135875702, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895474135875702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22305336594581604, + 0.10717715322971344, + 0.03789253160357475, + 0.039881519973278046, + 0.10800353437662125, + 0.04038359597325325, + 0.047200992703437805, + 0.04214171692728996, + 0.021048838272690773, + 0.3332167863845825 + ], + "credit_map": [ + 0.03163520619273186, + 0.04093952104449272, + 0.022503428161144257, + 0.01851208508014679, + 0.02529163286089897, + 0.03659714385867119, + 0.02499196119606495, + 0.022582391276955605, + 0.01299060508608818, + 0.02384118176996708, + 0.009366132318973541, + 0.008455747738480568, + 0.012618367560207844, + 0.017859064042568207, + 0.01225440576672554, + 0.009791651740670204, + 0.0051132552325725555, + 0.008625609800219536, + 0.004192433785647154, + 0.0031250943429768085, + 0.004142903722822666, + 0.006741577759385109, + 0.0028402376919984818, + 0.0031114176381379366, + 0.005240292754024267, + 0.009430965408682823, + 0.003313045483082533, + 0.003073484869673848, + 0.005623561330139637, + 0.007036639377474785, + 0.0030316647607833147, + 0.003131863661110401, + 0.013881796039640903, + 0.02461969666182995, + 0.017663342878222466, + 0.008109180256724358, + 0.009198086336255074, + 0.019001498818397522, + 0.007520779501646757, + 0.008009159006178379, + 0.006530598271638155, + 0.010352928191423416, + 0.004077523946762085, + 0.0033527296036481857, + 0.004058771301060915, + 0.007056301925331354, + 0.0026776802260428667, + 0.002277063438668847, + 0.004782236646860838, + 0.010383246466517448, + 0.00463566230610013, + 0.004944131709635258, + 0.0053770761005580425, + 0.0073082903400063515, + 0.004926234483718872, + 0.004844111856073141, + 0.006195924710482359, + 0.012555903755128384, + 0.003396865911781788, + 0.0036400535609573126, + 0.003320252289995551, + 0.008207933977246284, + 0.002917297650128603, + 0.0019074868177995086, + 0.0029634013772010803, + 0.004831009078770876, + 0.0022120685316622257, + 0.0015472926897928119, + 0.002690402092412114, + 0.003607900580391288, + 0.0013263043947517872, + 0.0018704583635553718, + 0.03966953605413437, + 0.10077613592147827, + 0.0271386057138443, + 0.019514258950948715, + 0.0400494709610939, + 0.07243902236223221, + 0.015605305321514606, + 0.018024468794465065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.07895474135875702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_15/index_tensor([9582, 9582])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_15/index_tensor([9582, 9582])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ec8b1dcc4a28a35f6b39d13e5ac7a0e945ee9aa0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_15/index_tensor([9582, 9582])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, filmed from a low angle, slowly pans forward, the overall image imbued with a distinctly animated style, characterized by clean lines and saturated colors. Deep within a vast, empty warehouse, a lone figure is absorbed in preparing dinner. This figure is a young man with tousled black hair, wearing a slightly oversized gray work T-shirt, his expression focused and determined. Before him lies a simple gas stove, atop which rests a massive, dark, traditional iron wok. He grips the wok's handle with both hands and, with almost frantic force and precise rhythm, violently lifts the wok upward. The vibrantly colored diced vegetables and meat within are instantly launched into the air, tracing a perfect arc before landing precisely back in the pot. A vibrant orange-red flame erupts from the stove, fiercely licking the pot's bottom, casting flickering shadows on the man's focused profile and the towering metal shelves surrounding him. These shelves were filled with countless unopened cardboard boxes and mysterious objects covered with canvas, casting huge and swaying shadows in the jumping firelight, forming a sharp contrast with the bottomless darkness in the distance, creating a strange atmosphere that was both lonely and full of vitality.", + "seed": 26223, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_15/index_9582_seed_26223_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0349102190695703, + "vlm_logprob": -0.0349102190695703, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.0349102190695703, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017806315794587135 + ], + "opposite_logprob_tokens": [ + -4.142806529998779 + ], + "logprob_reward": -0.017806315794587135, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017806315794587135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1614340841770172, + 0.06565820425748825, + 0.0817771852016449, + 0.05966823920607567, + 0.0667174831032753, + 0.2091469168663025, + 0.04918523132801056, + 0.1522616446018219, + 0.05902324244379997, + 0.09512779861688614 + ], + "credit_map": [ + 0.01919618435204029, + 0.022168589755892754, + 0.023766379803419113, + 0.02277311310172081, + 0.01748037151992321, + 0.020756570622324944, + 0.017390932887792587, + 0.017901943996548653, + 0.007858412340283394, + 0.009112995117902756, + 0.008543192408978939, + 0.009999466128647327, + 0.009399059228599072, + 0.007100509013980627, + 0.007146342657506466, + 0.006498231552541256, + 0.010724019259214401, + 0.012269005179405212, + 0.014482010155916214, + 0.009668570943176746, + 0.0074223740957677364, + 0.009283299557864666, + 0.009039982222020626, + 0.00888792797923088, + 0.0081569729372859, + 0.009383353404700756, + 0.008949311450123787, + 0.007502751890569925, + 0.006053767167031765, + 0.0071689835749566555, + 0.006582502741366625, + 0.005870596505701542, + 0.008191592060029507, + 0.00890996865928173, + 0.00959112960845232, + 0.00926197785884142, + 0.009696333669126034, + 0.007158983498811722, + 0.006892064120620489, + 0.0070154364220798016, + 0.02616959996521473, + 0.03586553409695625, + 0.03273221105337143, + 0.017854969948530197, + 0.016848519444465637, + 0.03405141457915306, + 0.016327377408742905, + 0.029297303408384323, + 0.005775439087301493, + 0.007006637752056122, + 0.006265717092901468, + 0.006081035826355219, + 0.005396033171564341, + 0.008547225967049599, + 0.004309435840696096, + 0.00580370519310236, + 0.03534473478794098, + 0.05015938729047775, + 0.00863798800855875, + 0.014920337125658989, + 0.010633799247443676, + 0.012961472384631634, + 0.01160042081028223, + 0.008003514260053635, + 0.005957963410764933, + 0.009374517947435379, + 0.012582264840602875, + 0.006730352528393269, + 0.006634521298110485, + 0.00770963728427887, + 0.005435164552181959, + 0.0045988187193870544, + 0.011138837784528732, + 0.013090504333376884, + 0.009894512593746185, + 0.015872422605752945, + 0.009095338173210621, + 0.011667586863040924, + 0.015237942337989807, + 0.009130648337304592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.017806315794587135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026357226073741913 + ], + "opposite_logprob_tokens": [ + -3.6513571739196777 + ], + "logprob_reward": -0.026357226073741913, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026357226073741913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09792924672365189, + 0.04201703146100044, + 0.0735761746764183, + 0.03203956410288811, + 0.027568571269512177, + 0.18467631936073303, + 0.13049045205116272, + 0.11334621906280518, + 0.04131181165575981, + 0.25704458355903625 + ], + "credit_map": [ + 0.010499387048184872, + 0.013914328999817371, + 0.0108834533020854, + 0.011168360710144043, + 0.012775164097547531, + 0.012308478355407715, + 0.014316840097308159, + 0.012063239701092243, + 0.004774915054440498, + 0.006168533116579056, + 0.0036655638832598925, + 0.005184774287045002, + 0.006546136923134327, + 0.004690366331487894, + 0.006155914627015591, + 0.00483083026483655, + 0.008164960891008377, + 0.010040296241641045, + 0.006902845576405525, + 0.012344284914433956, + 0.010245690122246742, + 0.008120850659906864, + 0.010506990365684032, + 0.00725026149302721, + 0.004251956008374691, + 0.005071765277534723, + 0.004101469647139311, + 0.003431336022913456, + 0.0031527401879429817, + 0.005112581420689821, + 0.0034299520775675774, + 0.0034877664875239134, + 0.003312899963930249, + 0.0038535194471478462, + 0.003948129713535309, + 0.0037997830659151077, + 0.004352763760834932, + 0.002588575007393956, + 0.003404111135751009, + 0.002308791736140847, + 0.01936235837638378, + 0.03173002600669861, + 0.01943962834775448, + 0.028444040566682816, + 0.025945371016860008, + 0.019009387120604515, + 0.02274893969297409, + 0.017996566370129585, + 0.014792747795581818, + 0.018395964056253433, + 0.014434734359383583, + 0.022819360718131065, + 0.015886733308434486, + 0.015583353117108345, + 0.015419995412230492, + 0.013157574459910393, + 0.013814672827720642, + 0.016967101022601128, + 0.014487026259303093, + 0.018733831122517586, + 0.012033602222800255, + 0.010573631152510643, + 0.017430191859602928, + 0.009306167252361774, + 0.004431135021150112, + 0.00533202663064003, + 0.005676462780684233, + 0.008524843491613865, + 0.003999581094831228, + 0.004206987097859383, + 0.004905876237899065, + 0.004234903492033482, + 0.0243365541100502, + 0.02728302776813507, + 0.030645452439785004, + 0.03692465275526047, + 0.03246760368347168, + 0.0317964181303978, + 0.04734954982995987, + 0.026241352781653404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.026357226073741913 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371204808354378 + ], + "opposite_logprob_tokens": [ + -3.408712148666382 + ], + "logprob_reward": -0.03371204808354378, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371204808354378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09045792371034622, + 0.04512685909867287, + 0.07772601395845413, + 0.022542085498571396, + 0.021735714748501778, + 0.10594982653856277, + 0.05663984641432762, + 0.2740277349948883, + 0.09043087065219879, + 0.21536310017108917 + ], + "credit_map": [ + 0.01057366468012333, + 0.015064925886690617, + 0.010432539507746696, + 0.008358112536370754, + 0.016054315492510796, + 0.01021976862102747, + 0.011263907887041569, + 0.008490697480738163, + 0.004728987347334623, + 0.00678748544305563, + 0.004561278503388166, + 0.006232411600649357, + 0.006583636626601219, + 0.004071125295013189, + 0.0066567943431437016, + 0.005505139473825693, + 0.009519493207335472, + 0.011324238032102585, + 0.012785233557224274, + 0.009624605067074299, + 0.008145062252879143, + 0.010042536072432995, + 0.009320554323494434, + 0.006964291445910931, + 0.002643448766320944, + 0.0035640369169414043, + 0.0031983989756554365, + 0.0028492375276982784, + 0.002318616723641753, + 0.0026067830622196198, + 0.0025428945664316416, + 0.002818670589476824, + 0.0029114193748682737, + 0.002755874302238226, + 0.0031287716701626778, + 0.0029286991339176893, + 0.0026942510157823563, + 0.0022780876606702805, + 0.0029002015944570303, + 0.0021384102292358875, + 0.013195199891924858, + 0.0168906357139349, + 0.00971461646258831, + 0.018362998962402344, + 0.013990700244903564, + 0.011093421839177608, + 0.014711324125528336, + 0.007990920916199684, + 0.007610402535647154, + 0.008476790972054005, + 0.00588868698105216, + 0.008926578797399998, + 0.007986783981323242, + 0.005971686914563179, + 0.006912876386195421, + 0.004866037983447313, + 0.06297991424798965, + 0.09019331634044647, + 0.01774017885327339, + 0.025737827643752098, + 0.020519418641924858, + 0.025176389142870903, + 0.018047712743282318, + 0.01363295316696167, + 0.008477394469082355, + 0.018345415592193604, + 0.013571541756391525, + 0.011365544982254505, + 0.008969886228442192, + 0.01568516157567501, + 0.009854273870587349, + 0.004161653108894825, + 0.020779788494110107, + 0.030922532081604004, + 0.024736419320106506, + 0.02808944694697857, + 0.026459170505404472, + 0.02674122340977192, + 0.03826235607266426, + 0.019372157752513885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.03371204808354378 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640188094228506 + ], + "opposite_logprob_tokens": [ + -4.882640361785889 + ], + "logprob_reward": -0.007640188094228506, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640188094228506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10131105780601501, + 0.03312638774514198, + 0.08163323998451233, + 0.02186228707432747, + 0.027934610843658447, + 0.16697591543197632, + 0.08439302444458008, + 0.15105105936527252, + 0.10057548433542252, + 0.23113693296909332 + ], + "credit_map": [ + 0.010031626559793949, + 0.014764432795345783, + 0.012421244755387306, + 0.011490102857351303, + 0.016295071691274643, + 0.011052237823605537, + 0.013556559570133686, + 0.011699776165187359, + 0.003171883523464203, + 0.004930121824145317, + 0.003051943378522992, + 0.004095936194062233, + 0.005371011793613434, + 0.0035168128088116646, + 0.004654386080801487, + 0.0043342942371964455, + 0.010292197577655315, + 0.011696061119437218, + 0.011067013256251812, + 0.01128482911735773, + 0.010656091384589672, + 0.009049689397215843, + 0.010289364494383335, + 0.007297991309314966, + 0.002639348153024912, + 0.0034846290946006775, + 0.0016389545053243637, + 0.0030179929453879595, + 0.002814691513776779, + 0.0029486354906111956, + 0.0036937943659722805, + 0.0016242420533671975, + 0.0029230816289782524, + 0.004156802780926228, + 0.0027293057646602392, + 0.004250104073435068, + 0.0038138225208967924, + 0.002871996257454157, + 0.004685345571488142, + 0.002504152012988925, + 0.02128785103559494, + 0.02759474329650402, + 0.01621817611157894, + 0.03482956439256668, + 0.019232813268899918, + 0.017077410593628883, + 0.015432190150022507, + 0.015303167514503002, + 0.011126857250928879, + 0.01253929641097784, + 0.009776891209185123, + 0.012916040606796741, + 0.011465827003121376, + 0.008303445763885975, + 0.01018127053976059, + 0.008083391934633255, + 0.027012869715690613, + 0.03876243159174919, + 0.015264404937624931, + 0.019689923152327538, + 0.013228625990450382, + 0.011609679087996483, + 0.01606632210314274, + 0.00941680558025837, + 0.008473810739815235, + 0.0237752515822649, + 0.013504785485565662, + 0.009816136211156845, + 0.007962568663060665, + 0.024214204400777817, + 0.006898390594869852, + 0.005930338986217976, + 0.02206125110387802, + 0.030363639816641808, + 0.0299533698707819, + 0.032383956015110016, + 0.019612565636634827, + 0.032707296311855316, + 0.03211596980690956, + 0.03193886578083038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.007640188094228506 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903531730175018 + ], + "opposite_logprob_tokens": [ + -2.4640352725982666 + ], + "logprob_reward": -0.08903531730175018, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903531730175018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1035265102982521, + 0.05305086448788643, + 0.081863634288311, + 0.029674319550395012, + 0.02893787994980812, + 0.12219765782356262, + 0.07760419696569443, + 0.22986599802970886, + 0.10647884011268616, + 0.1668001115322113 + ], + "credit_map": [ + 0.01356508582830429, + 0.016316406428813934, + 0.012521951459348202, + 0.012100976891815662, + 0.015736637637019157, + 0.008452200330793858, + 0.013813797384500504, + 0.01101946271955967, + 0.004716604948043823, + 0.007569627370685339, + 0.007563648745417595, + 0.005613360088318586, + 0.006409803405404091, + 0.008928723633289337, + 0.0069015477783977985, + 0.005347553174942732, + 0.008458488620817661, + 0.012048885226249695, + 0.009622817859053612, + 0.009512354619801044, + 0.007588602602481842, + 0.010574820451438427, + 0.010997848585247993, + 0.01305982656776905, + 0.0036298404447734356, + 0.004565993323922157, + 0.0036047319881618023, + 0.0037501410115510225, + 0.004258410073816776, + 0.0035108495503664017, + 0.0036136142443865538, + 0.0027407403104007244, + 0.003008844330906868, + 0.004446588456630707, + 0.003198627382516861, + 0.0038437882903963327, + 0.00563920009881258, + 0.0020566226448863745, + 0.00407423684373498, + 0.0026699767913669348, + 0.015605509281158447, + 0.02004864811897278, + 0.010514819994568825, + 0.02551981620490551, + 0.014887469820678234, + 0.01261933147907257, + 0.013337738811969757, + 0.009664339013397694, + 0.009299964644014835, + 0.011246977373957634, + 0.00831550918519497, + 0.01580098643898964, + 0.008430982939898968, + 0.007549659349024296, + 0.009895031340420246, + 0.007065092213451862, + 0.040373943746089935, + 0.05242614075541496, + 0.03321070969104767, + 0.021907195448875427, + 0.015788525342941284, + 0.026864049956202507, + 0.01825985126197338, + 0.02103559300303459, + 0.010418449528515339, + 0.014747344888746738, + 0.017041299492120743, + 0.016944505274295807, + 0.010013316757977009, + 0.014705031178891659, + 0.011248738504946232, + 0.011360171250998974, + 0.016190635040402412, + 0.01926024630665779, + 0.019491974264383316, + 0.024609709158539772, + 0.02161720022559166, + 0.019887804985046387, + 0.030244192108511925, + 0.015498343855142593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.08903531730175018 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, filmed from a low angle, slowly pans forward, the overall image imbued with a distinctly animated style, characterized by clean lines and saturated colors. Deep within a vast, empty warehouse, a lone figure is absorbed in preparing dinner. This figure is a young man with tousled black hair, wearing a slightly oversized gray work T-shirt, his expression focused and determined. Before him lies a simple gas stove, atop which rests a massive, dark, traditional iron wok. He grips the wok's handle with both hands and, with almost frantic force and precise rhythm, violently lifts the wok upward. The vibrantly colored diced vegetables and meat within are instantly launched into the air, tracing a perfect arc before landing precisely back in the pot. A vibrant orange-red flame erupts from the stove, fiercely licking the pot's bottom, casting flickering shadows on the man's focused profile and the towering metal shelves surrounding him. These shelves were filled with countless unopened cardboard boxes and mysterious objects covered with canvas, casting huge and swaying shadows in the jumping firelight, forming a sharp contrast with the bottomless darkness in the distance, creating a strange atmosphere that was both lonely and full of vitality.", + "seed": 26224, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_15/index_9582_seed_26224_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06650640852749348, + "vlm_logprob": -0.06650640852749348, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.06650640852749348, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16121433675289154 + ], + "opposite_logprob_tokens": [ + -1.9112143516540527 + ], + "logprob_reward": -0.16121433675289154, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16121433675289154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24585478007793427, + 0.0803597941994667, + 0.08905783295631409, + 0.10966725647449493, + 0.21855853497982025, + 0.102256178855896, + 0.04386179894208908, + 0.06380237638950348, + 0.020948613062500954, + 0.025632867589592934 + ], + "credit_map": [ + 0.028073398396372795, + 0.03161133453249931, + 0.029025740921497345, + 0.01946859061717987, + 0.028155770152807236, + 0.03718411177396774, + 0.025507433339953423, + 0.04682840779423714, + 0.010402314364910126, + 0.013395468704402447, + 0.007848297245800495, + 0.008197614923119545, + 0.012451290152966976, + 0.008162884041666985, + 0.00835379958152771, + 0.011548126116394997, + 0.00988822989165783, + 0.017582764849066734, + 0.010804918594658375, + 0.008683975785970688, + 0.012400196865200996, + 0.011895924806594849, + 0.010637296363711357, + 0.0071645169518888, + 0.010712571442127228, + 0.01957772672176361, + 0.012751564383506775, + 0.015851899981498718, + 0.01585565134882927, + 0.012334448285400867, + 0.008915814571082592, + 0.013667582534253597, + 0.02578716352581978, + 0.050906531512737274, + 0.014654617756605148, + 0.030586322769522667, + 0.036207228899002075, + 0.023161964491009712, + 0.020098505541682243, + 0.017156198620796204, + 0.012784418649971485, + 0.015086800791323185, + 0.008365633897483349, + 0.014224969781935215, + 0.013230419717729092, + 0.014209720306098461, + 0.011971145868301392, + 0.012383067049086094, + 0.0060251508839428425, + 0.0061275246553123, + 0.004614175297319889, + 0.00707221170887351, + 0.006357606500387192, + 0.004416944459080696, + 0.0037717693485319614, + 0.005476414691656828, + 0.008605641312897205, + 0.011475026607513428, + 0.00842618104070425, + 0.0071714031510055065, + 0.0071081784553825855, + 0.007733143400400877, + 0.004973866976797581, + 0.008308934047818184, + 0.002861546818166971, + 0.0029095096979290247, + 0.0033201505430042744, + 0.0021891442593187094, + 0.0025304502341896296, + 0.0022270632907748222, + 0.0027057735715061426, + 0.0022049753461033106, + 0.0029993748757988214, + 0.0031168200075626373, + 0.002474438864737749, + 0.004837090149521828, + 0.002943004248663783, + 0.0029774410650134087, + 0.003721715649589896, + 0.002562982030212879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.16121433675289154 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183792009949684 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183792009949684, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183792009949684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3695301413536072, + 0.15408214926719666, + 0.08944641798734665, + 0.04564937204122543, + 0.04053653031587601, + 0.10537721961736679, + 0.06325379014015198, + 0.043768446892499924, + 0.04348066449165344, + 0.04487522318959236 + ], + "credit_map": [ + 0.038096390664577484, + 0.055604517459869385, + 0.048743266612291336, + 0.03666439652442932, + 0.04235249385237694, + 0.04145628213882446, + 0.04087051376700401, + 0.06574229896068573, + 0.01760517619550228, + 0.020059581845998764, + 0.016138581559062004, + 0.017974426969885826, + 0.025657713413238525, + 0.02183550037443638, + 0.015299083665013313, + 0.019512087106704712, + 0.010909298434853554, + 0.015579809434711933, + 0.00908826757222414, + 0.010984339751303196, + 0.013901221565902233, + 0.011224077083170414, + 0.007657921873033047, + 0.010101488791406155, + 0.004694897681474686, + 0.007977072149515152, + 0.006365251261740923, + 0.005256010685116053, + 0.006141581106930971, + 0.005187422037124634, + 0.004723766818642616, + 0.0053033712320029736, + 0.005363651551306248, + 0.0059992228634655476, + 0.0037098615430295467, + 0.0045732236467301846, + 0.004322483204305172, + 0.0045290919952094555, + 0.003634336404502392, + 0.008404659107327461, + 0.014928543008863926, + 0.015321570448577404, + 0.010468512773513794, + 0.013803177513182163, + 0.009989442303776741, + 0.012752958573400974, + 0.009852663613855839, + 0.01826036348938942, + 0.009757116436958313, + 0.007569243665784597, + 0.0062522441148757935, + 0.007940170355141163, + 0.009580629877746105, + 0.007042274810373783, + 0.007057529874145985, + 0.008054586127400398, + 0.0060027227737009525, + 0.0074835061095654964, + 0.003823075909167528, + 0.004764856304973364, + 0.005677930079400539, + 0.005310594569891691, + 0.004125718493014574, + 0.006580040790140629, + 0.005050223786383867, + 0.007014977280050516, + 0.003161646192893386, + 0.006536964792758226, + 0.006468528416007757, + 0.005096235778182745, + 0.004065881948918104, + 0.006086215376853943, + 0.005476551596075296, + 0.006626930087804794, + 0.004862071014940739, + 0.005298211704939604, + 0.006020085420459509, + 0.00517948716878891, + 0.00599027331918478, + 0.005421618465334177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.018183792009949684 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043056320399045944 + ], + "opposite_logprob_tokens": [ + -3.1680562496185303 + ], + "logprob_reward": -0.043056320399045944, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043056320399045944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15700218081474304, + 0.05494935065507889, + 0.07266994565725327, + 0.11241830885410309, + 0.18055449426174164, + 0.11975041776895523, + 0.12586326897144318, + 0.10633834451436996, + 0.03892200067639351, + 0.0315316878259182 + ], + "credit_map": [ + 0.016684163361787796, + 0.02028917707502842, + 0.022036444395780563, + 0.012345701456069946, + 0.01802080310881138, + 0.021740004420280457, + 0.01681751012802124, + 0.029068375006318092, + 0.006330640986561775, + 0.007603333797305822, + 0.006203919183462858, + 0.007799278944730759, + 0.006899839732795954, + 0.0061055743135511875, + 0.005141502246260643, + 0.008865260519087315, + 0.007939259521663189, + 0.013912072405219078, + 0.007310586981475353, + 0.007709491066634655, + 0.010375510901212692, + 0.008812421932816505, + 0.008143573999404907, + 0.00846702791750431, + 0.013688752427697182, + 0.019186437129974365, + 0.017343156039714813, + 0.011621988378465176, + 0.012425192631781101, + 0.012170927599072456, + 0.008621875196695328, + 0.01735997572541237, + 0.017283741384744644, + 0.036280859261751175, + 0.012733621522784233, + 0.024534592404961586, + 0.026416197419166565, + 0.019247371703386307, + 0.020384157076478004, + 0.02367393672466278, + 0.014296849258244038, + 0.017044633626937866, + 0.009121527895331383, + 0.019226493313908577, + 0.017535872757434845, + 0.015274704433977604, + 0.013140001334249973, + 0.014110344462096691, + 0.014081954024732113, + 0.022048041224479675, + 0.01067806314677, + 0.02079458348453045, + 0.018623478710651398, + 0.014397640712559223, + 0.010606634430587292, + 0.014632881619036198, + 0.016465798020362854, + 0.01760326325893402, + 0.007649093866348267, + 0.015827622264623642, + 0.009753054939210415, + 0.015382286161184311, + 0.010522636584937572, + 0.013134588487446308, + 0.005503850989043713, + 0.006254431325942278, + 0.0040209051221609116, + 0.004572409670799971, + 0.00522246491163969, + 0.004349267575889826, + 0.004409273155033588, + 0.004589397925883532, + 0.00414274912327528, + 0.004645111039280891, + 0.0030675185844302177, + 0.0035901267547160387, + 0.004105234984308481, + 0.004073439631611109, + 0.004201841074973345, + 0.0037056650035083294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.043056320399045944 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785067290067673 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.009785067290067673, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785067290067673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32609909772872925, + 0.11580323427915573, + 0.0706716850399971, + 0.06673167645931244, + 0.05196184664964676, + 0.09868811815977097, + 0.0765146017074585, + 0.07091040909290314, + 0.08375705033540726, + 0.03886226937174797 + ], + "credit_map": [ + 0.04068839177489281, + 0.08164099603891373, + 0.03263450786471367, + 0.015933673828840256, + 0.05433052405714989, + 0.03988258168101311, + 0.017747540026903152, + 0.043240875005722046, + 0.013113345950841904, + 0.014261476695537567, + 0.012715401127934456, + 0.013216125778853893, + 0.017873425036668777, + 0.017544548958539963, + 0.010960989631712437, + 0.016117919236421585, + 0.007279355078935623, + 0.009862040169537067, + 0.010079976171255112, + 0.005242026876658201, + 0.007808448281139135, + 0.008251547813415527, + 0.006275956053286791, + 0.01587233506143093, + 0.007114131469279528, + 0.01697479374706745, + 0.005507655441761017, + 0.007171526085585356, + 0.01114314142614603, + 0.008780127391219139, + 0.00474791694432497, + 0.005292383022606373, + 0.007538789417594671, + 0.012596485204994678, + 0.002972331130877137, + 0.004626335576176643, + 0.006296772975474596, + 0.006256923079490662, + 0.0038865196984261274, + 0.007787690032273531, + 0.013843443244695663, + 0.014454271644353867, + 0.00787009671330452, + 0.01231022085994482, + 0.011615986004471779, + 0.012852187268435955, + 0.010410228744149208, + 0.015331682749092579, + 0.010706624016165733, + 0.008702388033270836, + 0.008915403857827187, + 0.009252361953258514, + 0.010666633024811745, + 0.009013538248836994, + 0.009836437180638313, + 0.009421218186616898, + 0.007941476069390774, + 0.01211591251194477, + 0.007637802045792341, + 0.010858594439923763, + 0.00810363981872797, + 0.008293514139950275, + 0.01005271915346384, + 0.0059067499823868275, + 0.010159661993384361, + 0.009695948101580143, + 0.013963855803012848, + 0.007093674503266811, + 0.009667031466960907, + 0.01077235210686922, + 0.011616796255111694, + 0.010787729173898697, + 0.00525063369423151, + 0.005424792412668467, + 0.004033950623124838, + 0.005774725694209337, + 0.005221224389970303, + 0.004987571854144335, + 0.004085738677531481, + 0.00408363388851285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.009785067290067673 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029252618551254 + ], + "opposite_logprob_tokens": [ + -2.350292444229126 + ], + "logprob_reward": -0.10029252618551254, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029252618551254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13373857736587524, + 0.0666520893573761, + 0.08440237492322922, + 0.13807041943073273, + 0.04878519847989082, + 0.113044373691082, + 0.08948986977338791, + 0.25504788756370544, + 0.020004015415906906, + 0.0507652647793293 + ], + "credit_map": [ + 0.01561375055462122, + 0.020868921652436256, + 0.015518761239945889, + 0.0123746357858181, + 0.016113659366965294, + 0.016156364232301712, + 0.013698533177375793, + 0.023393943905830383, + 0.008302893489599228, + 0.007343152072280645, + 0.010397769510746002, + 0.00662733381614089, + 0.0076883756555616856, + 0.007036297582089901, + 0.00525478832423687, + 0.01400147844105959, + 0.00885675847530365, + 0.011882662773132324, + 0.007460048887878656, + 0.010273306630551815, + 0.014042829163372517, + 0.01048087514936924, + 0.01181646715849638, + 0.009589429013431072, + 0.01444979477673769, + 0.022647587582468987, + 0.017473820596933365, + 0.014134367927908897, + 0.01659349352121353, + 0.014675147831439972, + 0.009930736385285854, + 0.028165467083454132, + 0.006395093165338039, + 0.007771956734359264, + 0.006027524825185537, + 0.005399639252573252, + 0.004929850343614817, + 0.005964836571365595, + 0.004568820353597403, + 0.007727479096502066, + 0.017227834090590477, + 0.018240291625261307, + 0.011733680963516235, + 0.015449095517396927, + 0.01026254054158926, + 0.012716195546090603, + 0.011218653060495853, + 0.016196073964238167, + 0.011390835978090763, + 0.009666590951383114, + 0.007838458754122257, + 0.010466049425303936, + 0.013434416614472866, + 0.014156613498926163, + 0.010726946406066418, + 0.01180995348840952, + 0.02966180630028248, + 0.04681030288338661, + 0.026735004037618637, + 0.03662523999810219, + 0.026309775188565254, + 0.02964719384908676, + 0.026846731081604958, + 0.03241182491183281, + 0.0026493975892663, + 0.003122272901237011, + 0.0021751585882157087, + 0.002331028925254941, + 0.0036920185666531324, + 0.0020228815264999866, + 0.0017399850767105818, + 0.0022712708450853825, + 0.00590042257681489, + 0.008962729945778847, + 0.005673393607139587, + 0.006097161211073399, + 0.0060022370889782906, + 0.005233628675341606, + 0.0059142084792256355, + 0.006981479469686747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.10029252618551254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, filmed from a low angle, slowly pans forward, the overall image imbued with a distinctly animated style, characterized by clean lines and saturated colors. Deep within a vast, empty warehouse, a lone figure is absorbed in preparing dinner. This figure is a young man with tousled black hair, wearing a slightly oversized gray work T-shirt, his expression focused and determined. Before him lies a simple gas stove, atop which rests a massive, dark, traditional iron wok. He grips the wok's handle with both hands and, with almost frantic force and precise rhythm, violently lifts the wok upward. The vibrantly colored diced vegetables and meat within are instantly launched into the air, tracing a perfect arc before landing precisely back in the pot. A vibrant orange-red flame erupts from the stove, fiercely licking the pot's bottom, casting flickering shadows on the man's focused profile and the towering metal shelves surrounding him. These shelves were filled with countless unopened cardboard boxes and mysterious objects covered with canvas, casting huge and swaying shadows in the jumping firelight, forming a sharp contrast with the bottomless darkness in the distance, creating a strange atmosphere that was both lonely and full of vitality.", + "seed": 26225, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_15/index_9582_seed_26225_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.28659039456397295, + "vlm_logprob": -0.28659039456397295, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.28659039456397295, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.137547492980957 + ], + "opposite_logprob_tokens": [ + -0.38754749298095703 + ], + "logprob_reward": -1.137547492980957, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.137547492980957, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.29145410656929016, + 0.10477598756551743, + 0.06350903958082199, + 0.05035857483744621, + 0.09584969282150269, + 0.059030864387750626, + 0.05363421514630318, + 0.07427049428224564, + 0.18257051706314087, + 0.024546509608626366 + ], + "credit_map": [ + 0.03806208446621895, + 0.06942901015281677, + 0.039613209664821625, + 0.021023083478212357, + 0.02798543870449066, + 0.043391309678554535, + 0.020457949489355087, + 0.03149202838540077, + 0.015602144412696362, + 0.029970217496156693, + 0.013207635842263699, + 0.006586892530322075, + 0.006614848971366882, + 0.018594222143292427, + 0.00477155577391386, + 0.00942847691476345, + 0.009963370859622955, + 0.013549246825277805, + 0.011059539392590523, + 0.004639769438654184, + 0.004547805991023779, + 0.01084062922745943, + 0.0031043223571032286, + 0.005804360378533602, + 0.005179634317755699, + 0.00804437231272459, + 0.0052973683923482895, + 0.0041124396957457066, + 0.010152203030884266, + 0.007374988403171301, + 0.005886156111955643, + 0.004311411175876856, + 0.012475957162678242, + 0.020196884870529175, + 0.015866748988628387, + 0.005299304146319628, + 0.008637402206659317, + 0.01747523993253708, + 0.005229586269706488, + 0.010668565519154072, + 0.008313323371112347, + 0.012510932981967926, + 0.009248318150639534, + 0.003956540487706661, + 0.004946579225361347, + 0.00983680970966816, + 0.003925560973584652, + 0.006292800884693861, + 0.009413231164216995, + 0.004882114008069038, + 0.007673707790672779, + 0.0036779262591153383, + 0.0058305817656219006, + 0.00812146533280611, + 0.008116042241454124, + 0.00591914402320981, + 0.012675449252128601, + 0.009516509249806404, + 0.01175925973802805, + 0.005507666617631912, + 0.010191261768341064, + 0.008399322628974915, + 0.007240802515298128, + 0.0089802211150527, + 0.028493160381913185, + 0.0257868655025959, + 0.038598231971263885, + 0.009395581670105457, + 0.017067881301045418, + 0.026749270036816597, + 0.00930581521242857, + 0.027173705399036407, + 0.003673115512356162, + 0.005319768097251654, + 0.00331426621414721, + 0.001928638550452888, + 0.0027250575367361307, + 0.0039291903376579285, + 0.001597434631548822, + 0.0020590382628142834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -1.137547492980957 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294130340218544 + ], + "opposite_logprob_tokens": [ + -3.77329421043396 + ], + "logprob_reward": -0.023294130340218544, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294130340218544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22255951166152954, + 0.07892321795225143, + 0.043346866965293884, + 0.1990485042333603, + 0.07719990611076355, + 0.0506041944026947, + 0.04790600761771202, + 0.10147849470376968, + 0.15198548138141632, + 0.026947740465402603 + ], + "credit_map": [ + 0.022992435842752457, + 0.042694006115198135, + 0.031201081350445747, + 0.019490886479616165, + 0.020160388201475143, + 0.03601989522576332, + 0.02324649691581726, + 0.026754314079880714, + 0.01121541578322649, + 0.019971583038568497, + 0.009839770384132862, + 0.00544973136857152, + 0.006164595950394869, + 0.013153846375644207, + 0.005984338000416756, + 0.007143939845263958, + 0.006775960326194763, + 0.00784304365515709, + 0.005627140868455172, + 0.003737845690920949, + 0.0049413712695240974, + 0.006809524726122618, + 0.002710208995267749, + 0.004901770036667585, + 0.01647334359586239, + 0.021010618656873703, + 0.02060772478580475, + 0.019275065511465073, + 0.04232972860336304, + 0.03118720091879368, + 0.028557781130075455, + 0.019607042893767357, + 0.008402843028306961, + 0.012675842270255089, + 0.009823698550462723, + 0.0068228780291974545, + 0.010194393806159496, + 0.01250922679901123, + 0.0071623921394348145, + 0.009608635678887367, + 0.0056107803247869015, + 0.00985331367701292, + 0.008485513739287853, + 0.004341007210314274, + 0.0054916818626224995, + 0.007869401946663857, + 0.004157352726906538, + 0.004795140121132135, + 0.009461913257837296, + 0.0052964165806770325, + 0.0051249125972390175, + 0.0038411584682762623, + 0.004936260636895895, + 0.006143280304968357, + 0.007279073819518089, + 0.00582299055531621, + 0.013129226863384247, + 0.012744824402034283, + 0.01564602553844452, + 0.008290409110486507, + 0.016642976552248, + 0.012303159572184086, + 0.010741215199232101, + 0.011980654671788216, + 0.026539336889982224, + 0.02146822027862072, + 0.03240150585770607, + 0.0077331457287073135, + 0.013639334589242935, + 0.02124406211078167, + 0.007634754758328199, + 0.021325131878256798, + 0.0031961530912667513, + 0.005382940638810396, + 0.004400302655994892, + 0.0016832250403240323, + 0.003386853262782097, + 0.004444370046257973, + 0.0016468423418700695, + 0.002807053504511714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.023294130340218544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904055505990982 + ], + "opposite_logprob_tokens": [ + -2.464040517807007 + ], + "logprob_reward": -0.08904055505990982, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904055505990982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23208355903625488, + 0.10707146674394608, + 0.044767096638679504, + 0.2157319039106369, + 0.0830635204911232, + 0.04745455086231232, + 0.08359453082084656, + 0.09380099177360535, + 0.05890890210866928, + 0.03352345898747444 + ], + "credit_map": [ + 0.03340646252036095, + 0.05104374885559082, + 0.034176792949438095, + 0.01668795384466648, + 0.024955378845334053, + 0.032205577939748764, + 0.01632874831557274, + 0.023278888314962387, + 0.014141981489956379, + 0.014454846270382404, + 0.014307082630693913, + 0.010552136227488518, + 0.017854219302535057, + 0.013761388137936592, + 0.012751735746860504, + 0.009248083457350731, + 0.004995178431272507, + 0.006774081848561764, + 0.008950487710535526, + 0.003722863271832466, + 0.004596063867211342, + 0.0067449756897985935, + 0.003783134277909994, + 0.005200312472879887, + 0.01936470903456211, + 0.02420433796942234, + 0.022158872336149216, + 0.019387023523449898, + 0.049171436578035355, + 0.03469179570674896, + 0.024540789425373077, + 0.022212928161025047, + 0.008400066755712032, + 0.009665720164775848, + 0.008281853049993515, + 0.008569668047130108, + 0.016968172043561935, + 0.011525521986186504, + 0.011188769713044167, + 0.008463747799396515, + 0.0054670581594109535, + 0.006956424098461866, + 0.005789346527308226, + 0.003968403209000826, + 0.00816384144127369, + 0.006462264806032181, + 0.0045056720264256, + 0.006141537334769964, + 0.016294004395604134, + 0.011153760366141796, + 0.010085265152156353, + 0.008628645911812782, + 0.009678984992206097, + 0.011412528343498707, + 0.008540217764675617, + 0.007801120635122061, + 0.012222280725836754, + 0.011013905517756939, + 0.008945423178374767, + 0.008539572358131409, + 0.020552074536681175, + 0.010977037250995636, + 0.010363267734646797, + 0.011187426745891571, + 0.0069879996590316296, + 0.009022746235132217, + 0.007969752885401249, + 0.005394812673330307, + 0.008213791996240616, + 0.007916687987744808, + 0.004800328053534031, + 0.00860278308391571, + 0.005651935003697872, + 0.005918830167502165, + 0.005233341362327337, + 0.0024095107801258564, + 0.0033727933187037706, + 0.004938499070703983, + 0.0028226803988218307, + 0.003175870282575488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.08904055505990982 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195776604115963 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.003195776604115963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195776604115963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2890692949295044, + 0.11188320815563202, + 0.0426810160279274, + 0.05535081401467323, + 0.056367821991443634, + 0.04551573842763901, + 0.0945001170039177, + 0.12408103793859482, + 0.13930442929267883, + 0.041246552020311356 + ], + "credit_map": [ + 0.040634289383888245, + 0.07301616668701172, + 0.034790731966495514, + 0.022407416254281998, + 0.029302332550287247, + 0.04302838072180748, + 0.01704414188861847, + 0.028845807537436485, + 0.01621955819427967, + 0.02722291462123394, + 0.012935033068060875, + 0.007541474886238575, + 0.012721110135316849, + 0.017419295385479927, + 0.008260879665613174, + 0.009562939405441284, + 0.005019006319344044, + 0.006595885846763849, + 0.006640063598752022, + 0.004082886036485434, + 0.005965405609458685, + 0.006697243545204401, + 0.003559676231816411, + 0.004120847210288048, + 0.00567892100661993, + 0.007652971427887678, + 0.0069716013967990875, + 0.005119640380144119, + 0.009843477047979832, + 0.008608660660684109, + 0.00542176328599453, + 0.0060537769459187984, + 0.007679141592234373, + 0.011425327509641647, + 0.008106185123324394, + 0.004109906032681465, + 0.005120293702930212, + 0.009714175947010517, + 0.0042900992557406425, + 0.005922688636928797, + 0.005561908707022667, + 0.00836602970957756, + 0.006584979593753815, + 0.003389239776879549, + 0.005781006533652544, + 0.0068311491049826145, + 0.0035092479083687067, + 0.005492177326232195, + 0.014936094172298908, + 0.008333981037139893, + 0.015197193250060081, + 0.008685354143381119, + 0.009123393334448338, + 0.012101820670068264, + 0.01466429140418768, + 0.01145798247307539, + 0.032407745718955994, + 0.013332689180970192, + 0.016987444832921028, + 0.0072338473983109, + 0.013007359579205513, + 0.015204420313239098, + 0.009926384314894676, + 0.01598113775253296, + 0.01716843992471695, + 0.02518937736749649, + 0.02527151256799698, + 0.00834388006478548, + 0.016305943951010704, + 0.02012518048286438, + 0.008509019389748573, + 0.01839105784893036, + 0.004696223884820938, + 0.007260490674525499, + 0.005852518603205681, + 0.0036869721952825785, + 0.0052098361775279045, + 0.0061004888266325, + 0.004105268511921167, + 0.004334749188274145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.003195776604115963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798740178346634 + ], + "opposite_logprob_tokens": [ + -1.804874062538147 + ], + "logprob_reward": -0.1798740178346634, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798740178346634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2279885858297348, + 0.08391173928976059, + 0.10680548846721649, + 0.0744851604104042, + 0.08161822706460953, + 0.04660357907414436, + 0.1174456775188446, + 0.1746235489845276, + 0.04050718620419502, + 0.04601079970598221 + ], + "credit_map": [ + 0.025439471006393433, + 0.03535798564553261, + 0.037184279412031174, + 0.024382274597883224, + 0.02724059857428074, + 0.03415839746594429, + 0.02492421679198742, + 0.019301362335681915, + 0.010224664583802223, + 0.02121920697391033, + 0.0066792722791433334, + 0.005823292303830385, + 0.01175131369382143, + 0.01312241330742836, + 0.009196741506457329, + 0.0058948383666574955, + 0.014267940074205399, + 0.02070571482181549, + 0.016551177948713303, + 0.007604838348925114, + 0.011963022872805595, + 0.01772070676088333, + 0.006382880732417107, + 0.011609201319515705, + 0.007273158058524132, + 0.011909428052604198, + 0.008106750436127186, + 0.008580709807574749, + 0.006292593665421009, + 0.011514975689351559, + 0.009638849645853043, + 0.011168694123625755, + 0.006635565310716629, + 0.011566255241632462, + 0.011784166097640991, + 0.007457556668668985, + 0.017891352996230125, + 0.011191045865416527, + 0.008220439776778221, + 0.006871842313557863, + 0.005976433865725994, + 0.006283876486122608, + 0.005187833681702614, + 0.0035732935648411512, + 0.00929408147931099, + 0.006965771783143282, + 0.0036233565770089626, + 0.005698930472135544, + 0.02530618943274021, + 0.017310453578829765, + 0.014112146571278572, + 0.009070733562111855, + 0.009303975850343704, + 0.01720009371638298, + 0.011983735486865044, + 0.013158347457647324, + 0.024484848603606224, + 0.022970374673604965, + 0.027925223112106323, + 0.012331051751971245, + 0.024298448115587234, + 0.023462720215320587, + 0.01814836636185646, + 0.02100250869989395, + 0.005208102986216545, + 0.0089290551841259, + 0.0050865765661001205, + 0.0033829936292022467, + 0.00439844373613596, + 0.0056998771615326405, + 0.0031335880048573017, + 0.004668549634516239, + 0.0067174616269767284, + 0.008075159043073654, + 0.006948682479560375, + 0.003769280156120658, + 0.004292615223675966, + 0.007153578568249941, + 0.004295039921998978, + 0.004758983384817839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.1798740178346634 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, filmed from a low angle, slowly pans forward, the overall image imbued with a distinctly animated style, characterized by clean lines and saturated colors. Deep within a vast, empty warehouse, a lone figure is absorbed in preparing dinner. This figure is a young man with tousled black hair, wearing a slightly oversized gray work T-shirt, his expression focused and determined. Before him lies a simple gas stove, atop which rests a massive, dark, traditional iron wok. He grips the wok's handle with both hands and, with almost frantic force and precise rhythm, violently lifts the wok upward. The vibrantly colored diced vegetables and meat within are instantly launched into the air, tracing a perfect arc before landing precisely back in the pot. A vibrant orange-red flame erupts from the stove, fiercely licking the pot's bottom, casting flickering shadows on the man's focused profile and the towering metal shelves surrounding him. These shelves were filled with countless unopened cardboard boxes and mysterious objects covered with canvas, casting huge and swaying shadows in the jumping firelight, forming a sharp contrast with the bottomless darkness in the distance, creating a strange atmosphere that was both lonely and full of vitality.", + "seed": 26226, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_15/index_9582_seed_26226_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0424543185159564, + "vlm_logprob": -0.0424543185159564, + "vlm_yes_no_margin": 3.774999976158142, + "vlm_token_credit": -0.0424543185159564, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021513784304261208 + ], + "opposite_logprob_tokens": [ + -3.8965137004852295 + ], + "logprob_reward": -0.021513784304261208, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021513784304261208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31932318210601807, + 0.13805744051933289, + 0.11738806217908859, + 0.08020695298910141, + 0.04257842153310776, + 0.049323663115501404, + 0.04083317890763283, + 0.09110482782125473, + 0.06149931252002716, + 0.05968496948480606 + ], + "credit_map": [ + 0.048204027116298676, + 0.04135410487651825, + 0.03600618615746498, + 0.014850749634206295, + 0.020196329802274704, + 0.058417439460754395, + 0.01700933650135994, + 0.08328504115343094, + 0.015466942451894283, + 0.015192244201898575, + 0.011328748427331448, + 0.008589464239776134, + 0.012962364591658115, + 0.02508489415049553, + 0.01064020674675703, + 0.038792580366134644, + 0.011566432192921638, + 0.023441681638360023, + 0.010206853039562702, + 0.005838694516569376, + 0.008457277901470661, + 0.022236254066228867, + 0.005308890715241432, + 0.030331982299685478, + 0.00733516551554203, + 0.01344574149698019, + 0.013580088503658772, + 0.006887710653245449, + 0.008944659493863583, + 0.013115051202476025, + 0.00630852160975337, + 0.01059001311659813, + 0.00449345400556922, + 0.008277352899312973, + 0.0032187490724027157, + 0.004695134237408638, + 0.005522306077182293, + 0.0073213521391153336, + 0.003201127052307129, + 0.005848952569067478, + 0.004696491174399853, + 0.007401963230222464, + 0.0038765498902648687, + 0.004953708034008741, + 0.005481830332428217, + 0.008393757976591587, + 0.00519178481772542, + 0.009327581152319908, + 0.006013544742017984, + 0.005702839232981205, + 0.0044374484568834305, + 0.0042620692402124405, + 0.004425062332302332, + 0.006286944728344679, + 0.003809093264862895, + 0.005896179471164942, + 0.011448907665908337, + 0.011389226652681828, + 0.007039851043373346, + 0.007070179097354412, + 0.013256027363240719, + 0.016820313408970833, + 0.008782045915722847, + 0.015298279002308846, + 0.00862029567360878, + 0.0074109709821641445, + 0.005463690962642431, + 0.004556648898869753, + 0.009369701147079468, + 0.009751669131219387, + 0.005452113226056099, + 0.010874221101403236, + 0.007975919172167778, + 0.006679473910480738, + 0.008056829683482647, + 0.004233560059219599, + 0.00895847287029028, + 0.009230908006429672, + 0.00436059245839715, + 0.010189215652644634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.021513784304261208 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583942532539368 + ], + "opposite_logprob_tokens": [ + -3.8955838680267334 + ], + "logprob_reward": -0.020583942532539368, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583942532539368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1338556557893753, + 0.2975858747959137, + 0.03844572231173515, + 0.10316663235425949, + 0.049707137048244476, + 0.09226051717996597, + 0.046922631561756134, + 0.042323384433984756, + 0.07664947956800461, + 0.11908295750617981 + ], + "credit_map": [ + 0.018219033256173134, + 0.017575299367308617, + 0.015409398823976517, + 0.012685460969805717, + 0.015642032027244568, + 0.019333817064762115, + 0.011202311143279076, + 0.02378830686211586, + 0.025851448997855186, + 0.034304868429899216, + 0.016429834067821503, + 0.015850553289055824, + 0.029619403183460236, + 0.05647220462560654, + 0.016459420323371887, + 0.1025981456041336, + 0.003932424820959568, + 0.004284003283828497, + 0.004946890752762556, + 0.0037722939159721136, + 0.004212541505694389, + 0.00669248728081584, + 0.0031892885453999043, + 0.007415792904794216, + 0.008484724909067154, + 0.014903044328093529, + 0.007979191839694977, + 0.01226919423788786, + 0.018387770280241966, + 0.017063211649656296, + 0.008916460908949375, + 0.015163029544055462, + 0.004774288274347782, + 0.009788934141397476, + 0.005286694038659334, + 0.004332458600401878, + 0.0079441973939538, + 0.007472853641957045, + 0.003768173512071371, + 0.0063395374454557896, + 0.00863422267138958, + 0.011801852844655514, + 0.00813369918614626, + 0.008402394130825996, + 0.007995443418622017, + 0.01689981482923031, + 0.008709555491805077, + 0.02168353646993637, + 0.006834487896412611, + 0.005253143142908812, + 0.006086747162044048, + 0.007708894554525614, + 0.004222668241709471, + 0.006784447468817234, + 0.004442555829882622, + 0.005589686334133148, + 0.0058234273456037045, + 0.004221899434924126, + 0.0035330038517713547, + 0.003929668106138706, + 0.004574465565383434, + 0.007422983646392822, + 0.0033281005453318357, + 0.009489831514656544, + 0.008385943248867989, + 0.009776654653251171, + 0.008294171653687954, + 0.005735005252063274, + 0.012214639224112034, + 0.012637297622859478, + 0.00804157368838787, + 0.011564189568161964, + 0.014594433829188347, + 0.013458043336868286, + 0.010772131383419037, + 0.009398573078215122, + 0.017002340406179428, + 0.018914664164185524, + 0.009942246600985527, + 0.025000520050525665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.020583942532539368 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023292500525712967 + ], + "opposite_logprob_tokens": [ + -3.7732925415039062 + ], + "logprob_reward": -0.023292500525712967, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023292500525712967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15914149582386017, + 0.14089477062225342, + 0.03197699785232544, + 0.06149735674262047, + 0.07988598942756653, + 0.12270260602235794, + 0.05580030754208565, + 0.10780920833349228, + 0.0819379985332489, + 0.15835322439670563 + ], + "credit_map": [ + 0.015574587509036064, + 0.021938089281320572, + 0.015556967817246914, + 0.011673138476908207, + 0.014607630670070648, + 0.025878990069031715, + 0.010924707166850567, + 0.04298737272620201, + 0.011531168594956398, + 0.017149638384580612, + 0.011649058200418949, + 0.007949338294565678, + 0.01581418141722679, + 0.025238176807761192, + 0.008355390280485153, + 0.04320782423019409, + 0.004611912649124861, + 0.003104542847722769, + 0.0030676827300339937, + 0.002029972616583109, + 0.0025945696979761124, + 0.005845020525157452, + 0.0027164712082594633, + 0.008006825111806393, + 0.006439159624278545, + 0.009551904164254665, + 0.010413626208901405, + 0.004942717496305704, + 0.006435272749513388, + 0.009435568004846573, + 0.005132682621479034, + 0.009146425873041153, + 0.009814965538680553, + 0.014449630863964558, + 0.005809355527162552, + 0.00897811446338892, + 0.011147079057991505, + 0.011693108826875687, + 0.006579617969691753, + 0.011414118111133575, + 0.01567167229950428, + 0.01149887964129448, + 0.011854946613311768, + 0.010012145154178143, + 0.009433760307729244, + 0.0210647601634264, + 0.010146954096853733, + 0.03301947936415672, + 0.008056693710386753, + 0.006576185114681721, + 0.0066824499517679214, + 0.005742909386754036, + 0.005261292215436697, + 0.009270457550883293, + 0.0056869713589549065, + 0.008523344062268734, + 0.012137098237872124, + 0.013320663012564182, + 0.009694666601717472, + 0.009764742106199265, + 0.018674025312066078, + 0.01574942097067833, + 0.011380709707736969, + 0.017087886109948158, + 0.010973045602440834, + 0.0057521662674844265, + 0.005991957616060972, + 0.007068248465657234, + 0.00946701131761074, + 0.014084997586905956, + 0.008356352336704731, + 0.020244214683771133, + 0.021156558766961098, + 0.013220944441854954, + 0.01773511804640293, + 0.01687704771757126, + 0.03050660900771618, + 0.024303879588842392, + 0.015103507786989212, + 0.019449569284915924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.023292500525712967 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098587669432163 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098587669432163, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098587669432163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1594315469264984, + 0.1823480874300003, + 0.03624327853322029, + 0.09167875349521637, + 0.07246863842010498, + 0.06986061483621597, + 0.05877945199608803, + 0.07576537132263184, + 0.09393903613090515, + 0.15948523581027985 + ], + "credit_map": [ + 0.017879173159599304, + 0.024510614573955536, + 0.016752183437347412, + 0.011798444204032421, + 0.020146746188402176, + 0.02529512159526348, + 0.014062143862247467, + 0.02898712269961834, + 0.02008197270333767, + 0.020440388470888138, + 0.012991596944630146, + 0.009665144607424736, + 0.013287659734487534, + 0.03537617623806, + 0.010167560540139675, + 0.06033758819103241, + 0.005160887259989977, + 0.003552674315869808, + 0.00408515240997076, + 0.0030473037622869015, + 0.003040545154362917, + 0.00618492066860199, + 0.0037708119489252567, + 0.007400981616228819, + 0.010528091341257095, + 0.011587983928620815, + 0.011298819445073605, + 0.009778385981917381, + 0.010151012800633907, + 0.014821990393102169, + 0.0056283725425601006, + 0.017884092405438423, + 0.009221384301781654, + 0.010936561971902847, + 0.0064139775931835175, + 0.006382978521287441, + 0.009126181714236736, + 0.010685427114367485, + 0.0078856460750103, + 0.011816486716270447, + 0.009929697029292583, + 0.007427976466715336, + 0.005740991793572903, + 0.004867832642048597, + 0.006807501893490553, + 0.01367172971367836, + 0.0042177713476121426, + 0.01719711534678936, + 0.007967228069901466, + 0.005246412940323353, + 0.00788883026689291, + 0.009098381735384464, + 0.0047947415150702, + 0.008628293871879578, + 0.005854136310517788, + 0.009301425889134407, + 0.00683521619066596, + 0.010939300060272217, + 0.0053026312962174416, + 0.008466453291475773, + 0.008865008130669594, + 0.013530626893043518, + 0.009325031191110611, + 0.01250110287219286, + 0.011136469431221485, + 0.008011492900550365, + 0.0057059829123318195, + 0.005456907674670219, + 0.012593397870659828, + 0.018682586029171944, + 0.008055352605879307, + 0.024296848103404045, + 0.024455538019537926, + 0.014485412277281284, + 0.02100178971886635, + 0.01663859933614731, + 0.023269381374120712, + 0.025531701743602753, + 0.011659440584480762, + 0.022443363443017006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.004098587669432163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast, empty warehouse with towering metal shelves filled with cardboard boxes and canvas-covered objects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with tousled black hair wearing a slightly oversized gray work T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man violently lifting a massive, dark, traditional iron wok with both hands, launching diced vegetables and meat into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image style distinctly animated with clean lines and saturated colors, featuring a vibrant orange-red flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot filmed from a low angle that slowly pans forward?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427827775478363 + ], + "opposite_logprob_tokens": [ + -2.017782688140869 + ], + "logprob_reward": -0.1427827775478363, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1427827775478363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15915042161941528, + 0.13485120236873627, + 0.06968618184328079, + 0.0597214549779892, + 0.11221934109926224, + 0.07363273948431015, + 0.17398753762245178, + 0.10741773247718811, + 0.05890064314007759, + 0.050432730466127396 + ], + "credit_map": [ + 0.01619829423725605, + 0.03143703565001488, + 0.012595899403095245, + 0.017604656517505646, + 0.01800505816936493, + 0.026672115549445152, + 0.017128612846136093, + 0.019508754834532738, + 0.012970003299415112, + 0.018231453374028206, + 0.009979279711842537, + 0.008289841935038567, + 0.009351701475679874, + 0.02515505626797676, + 0.007654903922230005, + 0.04321897774934769, + 0.005995958112180233, + 0.009553750045597553, + 0.005543149542063475, + 0.006819427013397217, + 0.00979628786444664, + 0.011991112492978573, + 0.006387827917933464, + 0.013598673045635223, + 0.006033772137016058, + 0.006866270210593939, + 0.006039612926542759, + 0.006872057914733887, + 0.004904860630631447, + 0.011237334460020065, + 0.0034910303074866533, + 0.014276517555117607, + 0.011433770880103111, + 0.023116474971175194, + 0.008561303839087486, + 0.013138892129063606, + 0.014326266013085842, + 0.01574564166367054, + 0.008980998769402504, + 0.01691598817706108, + 0.009367343969643116, + 0.010198050178587437, + 0.005308226682245731, + 0.004658717196434736, + 0.006903128232806921, + 0.014917937107384205, + 0.0065740784630179405, + 0.015705259516835213, + 0.03962332010269165, + 0.015983078628778458, + 0.02175194025039673, + 0.011092714965343475, + 0.013023164123296738, + 0.03022998385131359, + 0.01214517094194889, + 0.030138162896037102, + 0.014787722378969193, + 0.013483213260769844, + 0.00994094181805849, + 0.008422325365245342, + 0.016196580603718758, + 0.015056256204843521, + 0.007984167896211147, + 0.02154652588069439, + 0.006500760558992624, + 0.005858541466295719, + 0.0049950359389185905, + 0.005832422990351915, + 0.008812529034912586, + 0.009492781013250351, + 0.007570656016469002, + 0.009837917983531952, + 0.006353884004056454, + 0.004197992850095034, + 0.004364184569567442, + 0.0049912878312170506, + 0.00916995108127594, + 0.007896041497588158, + 0.004848521668463945, + 0.008610864169895649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.1427827775478363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_3/index_tensor([8935, 8935])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_3/index_tensor([8935, 8935])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..58468607d013b4dd337b15410a0e3043c3a518f3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_3/index_tensor([8935, 8935])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "majestic, slow-panning wide shot reveals an otherworldly underwater ballroom, immersed in crystal-clear azure waters. Towering white marble columns, adorned with ancient Greek-style carvings, stand majestically. Between them, almost ghostly, transparent jellyfish pulsate and float gracefully. Their long, translucent tentacles, adorned with bioluminescent algae that emit a soft blue-green glow, serve as an ever-shifting source of light in this serene world. Amidst these luminous creatures, formally dressed individuals—men in classic black tuxedos and women in opulent gowns in deep jewel tones like sapphire and emerald—slowly dance the waltz in near-weightlessness. They twirl and glide gracefully in a dreamy, slow rhythm, their silk and chiffon gowns unfurling and swaying like blossoming flowers in the gentle currents of the water. Above the dance floor, several giant sea turtles glided majestically and silently beneath the ballroom dome, their dark silhouettes casting shimmering, distorted reflections on the highly polished coral floor below, complementing the glow of the dancers and jellyfish to create a tranquil, surreal and magically romantic atmosphere.", + "seed": 26025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_3/index_8935_seed_26025_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07701114788651467, + "vlm_logprob": -0.07701114788651467, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.07701114788651467, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1828400194644928 + ], + "opposite_logprob_tokens": [ + -1.8078399896621704 + ], + "logprob_reward": -0.1828400194644928, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1828400194644928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5138950943946838, + 0.052581701427698135, + 0.04369143769145012, + 0.03635305538773537, + 0.1838787943124771, + 0.03325783461332321, + 0.052433647215366364, + 0.021346241235733032, + 0.03227356821298599, + 0.030288605019450188 + ], + "credit_map": [ + 0.18996630609035492, + 0.027933191508054733, + 0.03348970785737038, + 0.018740983679890633, + 0.026644881814718246, + 0.1759665161371231, + 0.014902986586093903, + 0.026250580325722694, + 0.007569229230284691, + 0.007037915289402008, + 0.006089850328862667, + 0.005777070764452219, + 0.006178748328238726, + 0.00939180701971054, + 0.005193643271923065, + 0.005343438591808081, + 0.005356545094400644, + 0.006401902064681053, + 0.006421834696084261, + 0.005438772030174732, + 0.006792278029024601, + 0.004759021569043398, + 0.0037574965972453356, + 0.004763591103255749, + 0.006598046515136957, + 0.006971669849008322, + 0.003999817650765181, + 0.0027628305833786726, + 0.0034738657996058464, + 0.004863104782998562, + 0.003339518094435334, + 0.004344205837696791, + 0.06244709715247154, + 0.008649652823805809, + 0.009165650233626366, + 0.006774553097784519, + 0.010682931169867516, + 0.06927557289600372, + 0.004469987470656633, + 0.012413362972438335, + 0.004822546150535345, + 0.003937081899493933, + 0.004334712866693735, + 0.002955503761768341, + 0.003996303770691156, + 0.003528069704771042, + 0.003275080118328333, + 0.006408535409718752, + 0.01080907043069601, + 0.004955734591931105, + 0.006495093926787376, + 0.006282208487391472, + 0.006026901304721832, + 0.007095248904079199, + 0.004963486455380917, + 0.005805904511362314, + 0.002820603083819151, + 0.0027198174502700567, + 0.0021109667140990496, + 0.0017803697846829891, + 0.0034399298019707203, + 0.0036885507870465517, + 0.0027863665018230677, + 0.001999636646360159, + 0.00436368165537715, + 0.0033935036044567823, + 0.003706397721543908, + 0.0023472479078918695, + 0.0048033553175628185, + 0.007580643519759178, + 0.0031213879119604826, + 0.0029573505744338036, + 0.0056087784469127655, + 0.004646189510822296, + 0.005847173277288675, + 0.0032951566390693188, + 0.0025456412695348263, + 0.0037025287747383118, + 0.002206523669883609, + 0.0024366152938455343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.1828400194644928 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497889593243599 + ], + "opposite_logprob_tokens": [ + -2.929978847503662 + ], + "logprob_reward": -0.05497889593243599, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497889593243599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5843239426612854, + 0.05033426731824875, + 0.13900503516197205, + 0.02596779353916645, + 0.10541924089193344, + 0.014018123038113117, + 0.01549152098596096, + 0.01746523752808571, + 0.02642703242599964, + 0.02154782973229885 + ], + "credit_map": [ + 0.22345246374607086, + 0.028345981612801552, + 0.03672979399561882, + 0.015972288325428963, + 0.02791813760995865, + 0.2066979557275772, + 0.013820460997521877, + 0.03138682246208191, + 0.005665061995387077, + 0.007739911321550608, + 0.005567975342273712, + 0.006244921591132879, + 0.00857522338628769, + 0.00633611623197794, + 0.005305150523781776, + 0.004899903200566769, + 0.015793591737747192, + 0.013702507130801678, + 0.014191004447638988, + 0.027739569544792175, + 0.01970478519797325, + 0.01526238489896059, + 0.019939616322517395, + 0.01267156470566988, + 0.0033442385029047728, + 0.004752345848828554, + 0.0026327779050916433, + 0.0026422671508044004, + 0.0027842933777719736, + 0.00471150828525424, + 0.002672438044101, + 0.0024279237259179354, + 0.02609877660870552, + 0.007770538330078125, + 0.009033177979290485, + 0.008315721526741982, + 0.008625869639217854, + 0.031746894121170044, + 0.006354004610329866, + 0.0074742562137544155, + 0.0020859481301158667, + 0.0018431744538247585, + 0.0015797597588971257, + 0.0016256340313702822, + 0.0020730074029415846, + 0.001909585203975439, + 0.0015070777153596282, + 0.0013939368072897196, + 0.002024300629273057, + 0.0016965561080724, + 0.0018138824962079525, + 0.0015560451429337263, + 0.0024669156409800053, + 0.0028257807716727257, + 0.0012525063939392567, + 0.001855533104389906, + 0.0026705453637987375, + 0.001880161464214325, + 0.0018957183929160237, + 0.0014448701404035091, + 0.002715590875595808, + 0.0028324758168309927, + 0.0020451631862670183, + 0.0019807121716439724, + 0.0039392621256411076, + 0.002652308437973261, + 0.002742020646110177, + 0.0016368773067370057, + 0.004847085103392601, + 0.005543013568967581, + 0.002720582066103816, + 0.002345884218811989, + 0.0035272203385829926, + 0.003803018480539322, + 0.003313409397378564, + 0.0017533297650516033, + 0.003267529420554638, + 0.0021164927165955305, + 0.002034281613305211, + 0.001732547883875668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.05497889593243599 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583825185894966 + ], + "opposite_logprob_tokens": [ + -3.8955838680267334 + ], + "logprob_reward": -0.020583825185894966, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583825185894966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5985808968544006, + 0.042862024158239365, + 0.04591771960258484, + 0.055857472121715546, + 0.03311998397111893, + 0.016429483890533447, + 0.030087919905781746, + 0.032688308507204056, + 0.08371525257825851, + 0.060740891844034195 + ], + "credit_map": [ + 0.22551152110099792, + 0.029861021786928177, + 0.038656167685985565, + 0.020706547424197197, + 0.03116236999630928, + 0.2059711217880249, + 0.015179186128079891, + 0.03153298795223236, + 0.00552984606474638, + 0.006026875227689743, + 0.004991620313376188, + 0.005062108393758535, + 0.006840675603598356, + 0.006771774031221867, + 0.003716635750606656, + 0.003922489006072283, + 0.004925398156046867, + 0.004835445899516344, + 0.004430829081684351, + 0.004259077366441488, + 0.007231256924569607, + 0.006692285183817148, + 0.007592699956148863, + 0.005950730759650469, + 0.009068384766578674, + 0.010158484801650047, + 0.005521571263670921, + 0.004855425097048283, + 0.004907179158180952, + 0.007866883650422096, + 0.005238933023065329, + 0.00824060756713152, + 0.004722124431282282, + 0.0034782628063112497, + 0.003996120300143957, + 0.003435485064983368, + 0.0043368833139538765, + 0.0043429178185760975, + 0.00307899690233171, + 0.005729192867875099, + 0.001989515032619238, + 0.0016797400312498212, + 0.0018580331234261394, + 0.0013251722557470202, + 0.0019650994800031185, + 0.002590566175058484, + 0.001407659612596035, + 0.0036136985290795565, + 0.005503811873495579, + 0.0024379047099500895, + 0.0033362023532390594, + 0.001930070691742003, + 0.005223905202001333, + 0.007393187377601862, + 0.002021987223997712, + 0.0022408512886613607, + 0.0036760750226676464, + 0.003650701604783535, + 0.004466441459953785, + 0.003679967485368252, + 0.004287863150238991, + 0.003943616524338722, + 0.004444506485015154, + 0.00453913863748312, + 0.01719517633318901, + 0.009632362984120846, + 0.009793498553335667, + 0.008334457874298096, + 0.015289628878235817, + 0.011309799738228321, + 0.007016523741185665, + 0.005143805406987667, + 0.008725779131054878, + 0.006556465290486813, + 0.008231955580413342, + 0.004128872416913509, + 0.011866201646625996, + 0.009139071218669415, + 0.00610309885814786, + 0.0059894477017223835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.020583825185894966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031528025865555 + ], + "opposite_logprob_tokens": [ + -2.3503153324127197 + ], + "logprob_reward": -0.10031528025865555, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031528025865555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21102748811244965, + 0.04195806384086609, + 0.15707255899906158, + 0.028719225898385048, + 0.03518577665090561, + 0.054688964039087296, + 0.03545146435499191, + 0.10288156569004059, + 0.25621360540390015, + 0.07680130749940872 + ], + "credit_map": [ + 0.0646100714802742, + 0.01564774289727211, + 0.017108339816331863, + 0.015584873966872692, + 0.014945706352591515, + 0.06152845919132233, + 0.009442772716283798, + 0.01215950958430767, + 0.0052461279556155205, + 0.004789232276380062, + 0.008317225612699986, + 0.0053366865031421185, + 0.004822915885597467, + 0.006442042533308268, + 0.002999519696459174, + 0.0040043131448328495, + 0.019339222460985184, + 0.01556752622127533, + 0.014571280218660831, + 0.0375712513923645, + 0.018155504018068314, + 0.018394015729427338, + 0.018362808972597122, + 0.015110938809812069, + 0.004976155702024698, + 0.005322725977748632, + 0.002531475853174925, + 0.003473706543445587, + 0.0027568412479013205, + 0.0029293233528733253, + 0.002796795452013612, + 0.0039322031661868095, + 0.004150725435465574, + 0.0040330528281629086, + 0.004856234882026911, + 0.003678069217130542, + 0.004374295007437468, + 0.005576249212026596, + 0.0031721133273094893, + 0.005345035810023546, + 0.008126313798129559, + 0.008361944928765297, + 0.007438689935952425, + 0.0048614321276545525, + 0.006487857084721327, + 0.006983066909015179, + 0.005189308896660805, + 0.007240350358188152, + 0.006289490032941103, + 0.003731959033757448, + 0.004781757481396198, + 0.0038031763397157192, + 0.0041261534206569195, + 0.006014232523739338, + 0.0031394718680530787, + 0.0035652248188853264, + 0.01210263092070818, + 0.010328605771064758, + 0.0118175707757473, + 0.010210108011960983, + 0.015156286768615246, + 0.015565697103738785, + 0.013577152974903584, + 0.014123516157269478, + 0.04933687299489975, + 0.027527084574103355, + 0.029317649081349373, + 0.025694428011775017, + 0.05352359265089035, + 0.030744923278689384, + 0.02431327849626541, + 0.015755753964185715, + 0.01526509691029787, + 0.010197552852332592, + 0.015993647277355194, + 0.005120890215039253, + 0.010677683167159557, + 0.007891268469393253, + 0.005698840599507093, + 0.005956323351711035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.10031528025865555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337718591094017 + ], + "opposite_logprob_tokens": [ + -3.6513376235961914 + ], + "logprob_reward": -0.026337718591094017, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337718591094017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4504934549331665, + 0.13478249311447144, + 0.16106924414634705, + 0.04835428670048714, + 0.06542380154132843, + 0.011627714149653912, + 0.048771873116493225, + 0.024711068719625473, + 0.024610307067632675, + 0.03015577420592308 + ], + "credit_map": [ + 0.16509555280208588, + 0.021751796826720238, + 0.03407686576247215, + 0.01543217059224844, + 0.022975215688347816, + 0.15395981073379517, + 0.012983083724975586, + 0.024218954145908356, + 0.02718910574913025, + 0.01752345822751522, + 0.014828940853476524, + 0.007688936777412891, + 0.016885703429579735, + 0.03473557531833649, + 0.006900954060256481, + 0.009029818698763847, + 0.020021982491016388, + 0.014677861705422401, + 0.017682775855064392, + 0.027660852298140526, + 0.018256666138768196, + 0.01847280189394951, + 0.027300002053380013, + 0.016996294260025024, + 0.00540660135447979, + 0.00910100620239973, + 0.0055403620935976505, + 0.007055273279547691, + 0.004107268992811441, + 0.006218939553946257, + 0.0051649087108671665, + 0.005759922321885824, + 0.020077303051948547, + 0.0029482655227184296, + 0.004316997714340687, + 0.003460100619122386, + 0.0051667215302586555, + 0.020780397579073906, + 0.0038893590681254864, + 0.004784659948199987, + 0.0013640928082168102, + 0.0018370901234447956, + 0.0012480621226131916, + 0.001564757782034576, + 0.0016131110023707151, + 0.0014465163694694638, + 0.0012914612889289856, + 0.0012626223033294082, + 0.010269972495734692, + 0.004267563112080097, + 0.007613037712872028, + 0.005559774115681648, + 0.005531041417270899, + 0.005693680141121149, + 0.004236009903252125, + 0.005600794684141874, + 0.004012078978121281, + 0.00296552455984056, + 0.003347711870446801, + 0.002418619580566883, + 0.002874418394640088, + 0.002911339048296213, + 0.003620333271101117, + 0.002561042085289955, + 0.0043282886035740376, + 0.0025334914680570364, + 0.0025540299247950315, + 0.0019425752107053995, + 0.00511557050049305, + 0.00404747948050499, + 0.0022774834651499987, + 0.0018113908590748906, + 0.007467256393283606, + 0.0023161310236901045, + 0.008240806870162487, + 0.0020112015772610903, + 0.003304446814581752, + 0.003389904974028468, + 0.0015890257200226188, + 0.001836999086663127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.026337718591094017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic, slow-panning wide shot reveals an otherworldly underwater ballroom, immersed in crystal-clear azure waters. Towering white marble columns, adorned with ancient Greek-style carvings, stand majestically. Between them, almost ghostly, transparent jellyfish pulsate and float gracefully. Their long, translucent tentacles, adorned with bioluminescent algae that emit a soft blue-green glow, serve as an ever-shifting source of light in this serene world. Amidst these luminous creatures, formally dressed individuals—men in classic black tuxedos and women in opulent gowns in deep jewel tones like sapphire and emerald—slowly dance the waltz in near-weightlessness. They twirl and glide gracefully in a dreamy, slow rhythm, their silk and chiffon gowns unfurling and swaying like blossoming flowers in the gentle currents of the water. Above the dance floor, several giant sea turtles glided majestically and silently beneath the ballroom dome, their dark silhouettes casting shimmering, distorted reflections on the highly polished coral floor below, complementing the glow of the dancers and jellyfish to create a tranquil, surreal and magically romantic atmosphere.", + "seed": 26026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_3/index_8935_seed_26026_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09548163264989853, + "vlm_logprob": -0.09548163264989853, + "vlm_yes_no_margin": 2.524999976158142, + "vlm_token_credit": -0.09548163264989853, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13346540927886963 + ], + "opposite_logprob_tokens": [ + -2.13346529006958 + ], + "logprob_reward": -0.13346540927886963, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13346540927886963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16343335807323456, + 0.060149114578962326, + 0.11479399353265762, + 0.13256794214248657, + 0.32086366415023804, + 0.09030419588088989, + 0.03844759240746498, + 0.01295333169400692, + 0.03273453563451767, + 0.03375227004289627 + ], + "credit_map": [ + 0.016904868185520172, + 0.02215762808918953, + 0.014530254527926445, + 0.021757276728749275, + 0.017591310665011406, + 0.028020823374390602, + 0.02032235823571682, + 0.022148843854665756, + 0.010107399895787239, + 0.011496828868985176, + 0.0077440193854272366, + 0.006720299832522869, + 0.005662316456437111, + 0.006754732690751553, + 0.005688214674592018, + 0.0059753055684268475, + 0.011700456961989403, + 0.023475227877497673, + 0.01860901340842247, + 0.010818611830472946, + 0.009202135726809502, + 0.01994481310248375, + 0.010710572823882103, + 0.010333159938454628, + 0.00691180769354105, + 0.047445885837078094, + 0.004247928038239479, + 0.046739038079977036, + 0.006766513921320438, + 0.010528797283768654, + 0.0060048471204936504, + 0.003923122305423021, + 0.03667478635907173, + 0.05205337330698967, + 0.039936285465955734, + 0.035357944667339325, + 0.0507316067814827, + 0.03377711400389671, + 0.03695693239569664, + 0.035375628620386124, + 0.009526951238512993, + 0.009361617267131805, + 0.015310518443584442, + 0.010244636796414852, + 0.01038014329969883, + 0.009964101016521454, + 0.015104729682207108, + 0.010411501862108707, + 0.005146777722984552, + 0.004728306084871292, + 0.004757165443152189, + 0.004559163935482502, + 0.0069776009768247604, + 0.0037338302936404943, + 0.0038118951488286257, + 0.004732850939035416, + 0.0013668180909007788, + 0.001974893268197775, + 0.0019148733699694276, + 0.0015474429819732904, + 0.001922220690175891, + 0.0017449992010369897, + 0.0011467136209830642, + 0.0013353711692616343, + 0.0031117007602006197, + 0.00396268954500556, + 0.004908919334411621, + 0.0030449542682617903, + 0.004032205790281296, + 0.008992662653326988, + 0.00239565153606236, + 0.0022857519797980785, + 0.003391390899196267, + 0.007795920129865408, + 0.003731413511559367, + 0.00277598574757576, + 0.0038907849229872227, + 0.0025261170230805874, + 0.003464768873527646, + 0.006175887305289507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.13346540927886963 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026372667402029037 + ], + "opposite_logprob_tokens": [ + -3.6513726711273193 + ], + "logprob_reward": -0.026372667402029037, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026372667402029037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08809832483530045, + 0.049278389662504196, + 0.08255395293235779, + 0.20782631635665894, + 0.037086229771375656, + 0.050553467124700546, + 0.02106465958058834, + 0.0358024425804615, + 0.025598052889108658, + 0.4021381735801697 + ], + "credit_map": [ + 0.008886667899787426, + 0.011647356674075127, + 0.009782442823052406, + 0.010612799786031246, + 0.010456094518303871, + 0.009799855761229992, + 0.009469452314078808, + 0.01744365319609642, + 0.005663417279720306, + 0.0068086301907896996, + 0.007214011158794165, + 0.009129977785050869, + 0.005010142456740141, + 0.0043846494518220425, + 0.005906958132982254, + 0.005160602740943432, + 0.007690838072448969, + 0.017387187108397484, + 0.012007074430584908, + 0.00793349090963602, + 0.006954836659133434, + 0.01615052856504917, + 0.006729830056428909, + 0.00770016573369503, + 0.01016811653971672, + 0.07509248703718185, + 0.005643353331834078, + 0.07613549381494522, + 0.010251323692500591, + 0.017176957800984383, + 0.008152325637638569, + 0.005206270143389702, + 0.0034156031906604767, + 0.006881455425173044, + 0.004312797449529171, + 0.00386926275677979, + 0.006293812301009893, + 0.004292558878660202, + 0.00429186038672924, + 0.0037288819439709187, + 0.0062550753355026245, + 0.008343776687979698, + 0.010080169886350632, + 0.006536511704325676, + 0.0044533987529575825, + 0.0044027348048985004, + 0.004423624835908413, + 0.006058174185454845, + 0.0016009638784453273, + 0.00335162621922791, + 0.003204604145139456, + 0.0023029912263154984, + 0.0024638876784592867, + 0.002516312524676323, + 0.0017416193149983883, + 0.003882654244080186, + 0.003656713990494609, + 0.004882258363068104, + 0.006376136559993029, + 0.005092088133096695, + 0.004424629732966423, + 0.003758166218176484, + 0.0038976301439106464, + 0.003714820835739374, + 0.0020529702305793762, + 0.0029011801816523075, + 0.0044775065034627914, + 0.002009536372497678, + 0.0032371485140174627, + 0.006304534152150154, + 0.002681364305317402, + 0.0019338112324476242, + 0.03669201582670212, + 0.12781956791877747, + 0.02203322760760784, + 0.015990665182471275, + 0.03059082292020321, + 0.01899988017976284, + 0.017899202182888985, + 0.1321127861738205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.026372667402029037 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865739122033119 + ], + "opposite_logprob_tokens": [ + -3.0486574172973633 + ], + "logprob_reward": -0.04865739122033119, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865739122033119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09815425425767899, + 0.045525822788476944, + 0.4945332705974579, + 0.12040634453296661, + 0.030478721484541893, + 0.04584566131234169, + 0.010339034721255302, + 0.01944144442677498, + 0.021413622424006462, + 0.1138618215918541 + ], + "credit_map": [ + 0.007845570333302021, + 0.01195849385112524, + 0.012033786624670029, + 0.013066226616501808, + 0.01256523933261633, + 0.012479318305850029, + 0.013106482103466988, + 0.01509914267808199, + 0.007302980404347181, + 0.008736548945307732, + 0.003697218606248498, + 0.007981233298778534, + 0.004534226842224598, + 0.004972563590854406, + 0.004656635690480471, + 0.003644420998170972, + 0.04791406914591789, + 0.09911832213401794, + 0.07366029918193817, + 0.048167333006858826, + 0.042237844318151474, + 0.09272002428770065, + 0.049218952655792236, + 0.04149644076824188, + 0.0057592433877289295, + 0.04451807588338852, + 0.0028159746434539557, + 0.04388575628399849, + 0.005968219600617886, + 0.009950035251677036, + 0.004773854278028011, + 0.002735200570896268, + 0.0033832218032330275, + 0.004242350813001394, + 0.003723371773958206, + 0.0032112724147737026, + 0.00518047995865345, + 0.003129632445052266, + 0.003907601814717054, + 0.003700793720781803, + 0.004414588678628206, + 0.005477804224938154, + 0.008458912372589111, + 0.004811306018382311, + 0.005197698250412941, + 0.004356118850409985, + 0.007589058019220829, + 0.005540183279663324, + 0.0016581447562202811, + 0.0014267894439399242, + 0.0013291845098137856, + 0.0010466958628967404, + 0.0009650574065744877, + 0.00142396439332515, + 0.0009708493598736823, + 0.0015183500945568085, + 0.001808467786759138, + 0.002902647713199258, + 0.0036619259044528008, + 0.00287607847712934, + 0.0025646723806858063, + 0.0018762795953080058, + 0.0019304035231471062, + 0.0018209711415693164, + 0.0019488821271806955, + 0.0030228348914533854, + 0.002951243193820119, + 0.001947159180417657, + 0.0026494895573705435, + 0.005164209753274918, + 0.002078874735161662, + 0.0016509328270331025, + 0.010145762003958225, + 0.035407524555921555, + 0.0065724970772862434, + 0.0044619496911764145, + 0.008631938137114048, + 0.005567312240600586, + 0.004816154949367046, + 0.03825869411230087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.04865739122033119 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990867793560028 + ], + "opposite_logprob_tokens": [ + -1.8049086332321167 + ], + "logprob_reward": -0.17990867793560028, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990867793560028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09799939393997192, + 0.027012759819626808, + 0.077254518866539, + 0.32457295060157776, + 0.23024074733257294, + 0.05053712800145149, + 0.016635430976748466, + 0.025350073352456093, + 0.0963369682431221, + 0.054060060530900955 + ], + "credit_map": [ + 0.009228982031345367, + 0.011534908786416054, + 0.010378435254096985, + 0.012087161652743816, + 0.013744990341365337, + 0.011834247037768364, + 0.012528738938272, + 0.016661932691931725, + 0.0026637818664312363, + 0.005742923356592655, + 0.0020728493109345436, + 0.0057060932740569115, + 0.0025401650927960873, + 0.002511255443096161, + 0.0033308519050478935, + 0.0024448398035019636, + 0.0077715544030070305, + 0.0145020242780447, + 0.010996980592608452, + 0.008183980360627174, + 0.0072814649902284145, + 0.014756494201719761, + 0.00741279823705554, + 0.006349219474941492, + 0.01585633121430874, + 0.11925198137760162, + 0.007130410522222519, + 0.11986558139324188, + 0.01572995074093342, + 0.026885734871029854, + 0.01261263620108366, + 0.0072403280064463615, + 0.025208426639437675, + 0.040977973490953445, + 0.027277130633592606, + 0.026024140417575836, + 0.035833489149808884, + 0.0242746751755476, + 0.02572462148964405, + 0.024920295923948288, + 0.005253338720649481, + 0.004908062983304262, + 0.008723887614905834, + 0.0055541954934597015, + 0.005619077477604151, + 0.005813247989863157, + 0.007802742533385754, + 0.006862574256956577, + 0.0022004691418260336, + 0.002256981562823057, + 0.002354297088459134, + 0.002801604103296995, + 0.0014069759054109454, + 0.0017838962376117706, + 0.0019648498855531216, + 0.001866357633844018, + 0.002069484908133745, + 0.0030494232196360826, + 0.00506486464291811, + 0.0038520973175764084, + 0.003531086491420865, + 0.0025203574914485216, + 0.0026814991142600775, + 0.002581260399892926, + 0.006355596240609884, + 0.008891351521015167, + 0.016889307647943497, + 0.006043091416358948, + 0.013151165097951889, + 0.03039432503283024, + 0.008983464911580086, + 0.00562865799292922, + 0.004578050691634417, + 0.015468762256205082, + 0.004110740032047033, + 0.003326254663988948, + 0.005234488286077976, + 0.003039910923689604, + 0.005467969458550215, + 0.012833886779844761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.17990867793560028 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890040174126625 + ], + "opposite_logprob_tokens": [ + -2.4640040397644043 + ], + "logprob_reward": -0.0890040174126625, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890040174126625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06562399864196777, + 0.02860497683286667, + 0.140536367893219, + 0.3750361204147339, + 0.23513755202293396, + 0.04707377776503563, + 0.05091746523976326, + 0.014541268348693848, + 0.01900205761194229, + 0.02352650836110115 + ], + "credit_map": [ + 0.006153560243546963, + 0.00943618081510067, + 0.009266635403037071, + 0.010640576481819153, + 0.009203325025737286, + 0.007655972149223089, + 0.006672271061688662, + 0.0065954686142504215, + 0.004231633152812719, + 0.004147252067923546, + 0.0031608114950358868, + 0.004157259128987789, + 0.002853608224540949, + 0.003926408942788839, + 0.003235669806599617, + 0.0028923333156853914, + 0.013865635730326176, + 0.02680373750627041, + 0.02160072885453701, + 0.014436769299209118, + 0.012729703448712826, + 0.025814957916736603, + 0.013899781741201878, + 0.011385026387870312, + 0.01837277039885521, + 0.13766999542713165, + 0.008586074225604534, + 0.13814669847488403, + 0.0186920166015625, + 0.03124633990228176, + 0.014073789119720459, + 0.008248374797403812, + 0.027346208691596985, + 0.036087747663259506, + 0.03225428983569145, + 0.02712424471974373, + 0.03873233124613762, + 0.022367235273122787, + 0.026367152109742165, + 0.024858325719833374, + 0.005276530981063843, + 0.005792680196464062, + 0.0081301499158144, + 0.004772278480231762, + 0.00525759719312191, + 0.00536772096529603, + 0.005702098831534386, + 0.006774716079235077, + 0.004789526574313641, + 0.007479799911379814, + 0.008438874036073685, + 0.005705676041543484, + 0.005256056785583496, + 0.0057813189923763275, + 0.004221397452056408, + 0.009244806133210659, + 0.001303801778703928, + 0.0023880996741354465, + 0.002675445517525077, + 0.0026947716251015663, + 0.001540773781016469, + 0.0011358713964000344, + 0.0013772768434137106, + 0.001425226335413754, + 0.001907792640849948, + 0.0025383864995092154, + 0.0036875340156257153, + 0.002656926168128848, + 0.0020141343120485544, + 0.0017259953310713172, + 0.002968747867271304, + 0.0015025403117761016, + 0.002427683211863041, + 0.0031038185115903616, + 0.004686616361141205, + 0.0016252072528004646, + 0.0018987477524206042, + 0.0015509738586843014, + 0.0022353583481162786, + 0.005998099688440561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.0890040174126625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic, slow-panning wide shot reveals an otherworldly underwater ballroom, immersed in crystal-clear azure waters. Towering white marble columns, adorned with ancient Greek-style carvings, stand majestically. Between them, almost ghostly, transparent jellyfish pulsate and float gracefully. Their long, translucent tentacles, adorned with bioluminescent algae that emit a soft blue-green glow, serve as an ever-shifting source of light in this serene world. Amidst these luminous creatures, formally dressed individuals—men in classic black tuxedos and women in opulent gowns in deep jewel tones like sapphire and emerald—slowly dance the waltz in near-weightlessness. They twirl and glide gracefully in a dreamy, slow rhythm, their silk and chiffon gowns unfurling and swaying like blossoming flowers in the gentle currents of the water. Above the dance floor, several giant sea turtles glided majestically and silently beneath the ballroom dome, their dark silhouettes casting shimmering, distorted reflections on the highly polished coral floor below, complementing the glow of the dancers and jellyfish to create a tranquil, surreal and magically romantic atmosphere.", + "seed": 26027, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_3/index_8935_seed_26027_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08821550942957401, + "vlm_logprob": -0.08821550942957401, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.08821550942957401, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2049461454153061 + ], + "opposite_logprob_tokens": [ + -1.7049461603164673 + ], + "logprob_reward": -0.2049461454153061, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2049461454153061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1428050547838211, + 0.13655835390090942, + 0.33178582787513733, + 0.06709445267915726, + 0.14770913124084473, + 0.034684162586927414, + 0.03655175119638443, + 0.033547982573509216, + 0.02106616459786892, + 0.04819712042808533 + ], + "credit_map": [ + 0.016559874638915062, + 0.019579283893108368, + 0.014081403613090515, + 0.016211049631237984, + 0.021614519879221916, + 0.019709674641489983, + 0.012791248969733715, + 0.022258015349507332, + 0.015460813418030739, + 0.025468967854976654, + 0.010519170202314854, + 0.014871853403747082, + 0.021366238594055176, + 0.02019103802740574, + 0.01403842680156231, + 0.014641856774687767, + 0.03140554577112198, + 0.0461132787168026, + 0.0389682874083519, + 0.055871713906526566, + 0.037212058901786804, + 0.04465598240494728, + 0.044568177312612534, + 0.032990794628858566, + 0.004348116926848888, + 0.010683180764317513, + 0.0033350582234561443, + 0.003922498784959316, + 0.006716462317854166, + 0.020856210961937904, + 0.00437782471999526, + 0.012855105102062225, + 0.014881707727909088, + 0.018846360966563225, + 0.011838964186608791, + 0.02183150313794613, + 0.01605393923819065, + 0.026504384353756905, + 0.0193790215998888, + 0.018373269587755203, + 0.0033206839580088854, + 0.0028677063528448343, + 0.0025428577791899443, + 0.0036597216967493296, + 0.0037275857757776976, + 0.009157772175967693, + 0.003796444507315755, + 0.0056113917380571365, + 0.003502540523186326, + 0.00447395583614707, + 0.0034625865519046783, + 0.00447511812672019, + 0.006342163309454918, + 0.005559132434427738, + 0.004905475303530693, + 0.0038307830691337585, + 0.0036580555606633425, + 0.004860501270741224, + 0.002519942820072174, + 0.003954538144171238, + 0.006972284521907568, + 0.00423782505095005, + 0.0034895187709480524, + 0.0038553178310394287, + 0.0019443512428551912, + 0.0018486358458176255, + 0.0016164380358532071, + 0.002915871562436223, + 0.0037481505423784256, + 0.0029004395473748446, + 0.0036179095041006804, + 0.002474370878189802, + 0.004389412235468626, + 0.004096158780157566, + 0.0031510351691395044, + 0.007588140666484833, + 0.0038495042826980352, + 0.0120967086404562, + 0.0036013710778206587, + 0.009424793533980846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.2049461454153061 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029829679057002068 + ], + "opposite_logprob_tokens": [ + -3.529829740524292 + ], + "logprob_reward": -0.029829679057002068, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029829679057002068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13173319399356842, + 0.07314454764127731, + 0.08945805579423904, + 0.36918169260025024, + 0.10540251433849335, + 0.04270942136645317, + 0.05088179558515549, + 0.03298787400126457, + 0.029536409303545952, + 0.07496452331542969 + ], + "credit_map": [ + 0.01440518256276846, + 0.015549982897937298, + 0.016015825793147087, + 0.020915977656841278, + 0.014047951437532902, + 0.018099578097462654, + 0.017568232491612434, + 0.015130456537008286, + 0.007921473123133183, + 0.00798098836094141, + 0.006347391754388809, + 0.007539453450590372, + 0.014589418657124043, + 0.00987872201949358, + 0.012022954411804676, + 0.006864150054752827, + 0.007601891178637743, + 0.009340339340269566, + 0.007375097367912531, + 0.007492667529731989, + 0.017687834799289703, + 0.010771960020065308, + 0.018566468730568886, + 0.010621796362102032, + 0.01848217472434044, + 0.030581993982195854, + 0.006821718066930771, + 0.008620202541351318, + 0.039841581135988235, + 0.1550208479166031, + 0.01250192616134882, + 0.09731122106313705, + 0.01057288609445095, + 0.015045743435621262, + 0.008757415227591991, + 0.013344420120120049, + 0.014948759227991104, + 0.01888665370643139, + 0.012729242444038391, + 0.011117394082248211, + 0.00546350609511137, + 0.0035511550959199667, + 0.004096310120075941, + 0.007020852062851191, + 0.005909223575145006, + 0.0073470608331263065, + 0.0052323476411402225, + 0.004088966175913811, + 0.003776648547500372, + 0.005326236132532358, + 0.003935609944164753, + 0.005179125815629959, + 0.008990403264760971, + 0.009309573099017143, + 0.008529537357389927, + 0.005834661424160004, + 0.003718162886798382, + 0.005615700967609882, + 0.002943214727565646, + 0.0036816217470914125, + 0.005407403223216534, + 0.005097231362015009, + 0.0036313964519649744, + 0.0028931444976478815, + 0.0027420385740697384, + 0.0035170200280845165, + 0.0029624912422150373, + 0.0032916883938014507, + 0.00441086245700717, + 0.004472975153476, + 0.0050858985632658005, + 0.003053433494642377, + 0.007557048462331295, + 0.004436456598341465, + 0.0037860882002860308, + 0.007297162432223558, + 0.006075584329664707, + 0.023453492671251297, + 0.005628432147204876, + 0.016730256378650665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.029829679057002068 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369925171136856 + ], + "opposite_logprob_tokens": [ + -3.4086992740631104 + ], + "logprob_reward": -0.03369925171136856, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369925171136856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13190685212612152, + 0.07108721137046814, + 0.08775055408477783, + 0.10155914723873138, + 0.1439133584499359, + 0.060147788375616074, + 0.059716660529375076, + 0.05802193284034729, + 0.0735049918293953, + 0.2123914510011673 + ], + "credit_map": [ + 0.02556181512773037, + 0.0164654403924942, + 0.014735029079020023, + 0.014246968552470207, + 0.0175718292593956, + 0.016204074025154114, + 0.014026883989572525, + 0.013094821944832802, + 0.008460331708192825, + 0.00735542131587863, + 0.008617687039077282, + 0.008162477985024452, + 0.009934790432453156, + 0.01075474638491869, + 0.009309381246566772, + 0.008492376655340195, + 0.012018091976642609, + 0.013093139044940472, + 0.008915477432310581, + 0.009653803892433643, + 0.011339067481458187, + 0.013723901472985744, + 0.00834229588508606, + 0.010664776898920536, + 0.005845060106366873, + 0.011166699230670929, + 0.0034812826197594404, + 0.00487840361893177, + 0.010992566123604774, + 0.03770246356725693, + 0.004644639790058136, + 0.022848034277558327, + 0.01472865603864193, + 0.020794931799173355, + 0.01717264950275421, + 0.01774156093597412, + 0.017209796234965324, + 0.02289258874952793, + 0.019839493557810783, + 0.013533681631088257, + 0.006090917158871889, + 0.005435438361018896, + 0.004445655737072229, + 0.00615546153858304, + 0.007609983440488577, + 0.01647380366921425, + 0.005106874275952578, + 0.008829653263092041, + 0.005564307793974876, + 0.009704413823783398, + 0.0065426304936409, + 0.007516307756304741, + 0.008884537033736706, + 0.007714593783020973, + 0.006988373585045338, + 0.006801500916481018, + 0.007003501523286104, + 0.008566383272409439, + 0.005331853404641151, + 0.007616566028445959, + 0.010221851989626884, + 0.008228288032114506, + 0.0055984267964959145, + 0.005455065052956343, + 0.006682927720248699, + 0.007938184775412083, + 0.0051839761435985565, + 0.010899160988628864, + 0.01509612426161766, + 0.008509566076099873, + 0.011724873445928097, + 0.007470176089555025, + 0.02225295826792717, + 0.012944130226969719, + 0.010288543067872524, + 0.01764872297644615, + 0.013309993781149387, + 0.07031720131635666, + 0.013216685503721237, + 0.052413210272789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.03369925171136856 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428101658821106 + ], + "opposite_logprob_tokens": [ + -2.017810106277466 + ], + "logprob_reward": -0.1428101658821106, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428101658821106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09288628399372101, + 0.052228689193725586, + 0.11677948385477066, + 0.36115431785583496, + 0.17561179399490356, + 0.04234214127063751, + 0.023138338699936867, + 0.025374438613653183, + 0.03663826733827591, + 0.07384629547595978 + ], + "credit_map": [ + 0.009501797147095203, + 0.013030174188315868, + 0.009754111059010029, + 0.010387648828327656, + 0.01236774679273367, + 0.009583747945725918, + 0.013533541932702065, + 0.014727498404681683, + 0.007350514642894268, + 0.010943794623017311, + 0.00340617005713284, + 0.004514625761657953, + 0.007291123736649752, + 0.007652972359210253, + 0.006059417966753244, + 0.005010065622627735, + 0.011350017972290516, + 0.016655027866363525, + 0.00978627149015665, + 0.022901073098182678, + 0.01353358756750822, + 0.014315539970993996, + 0.017377452924847603, + 0.010860499925911427, + 0.017330888658761978, + 0.031326401978731155, + 0.007963021285831928, + 0.010174690745770931, + 0.0380372554063797, + 0.14950378239154816, + 0.01269080676138401, + 0.09412741661071777, + 0.015343801118433475, + 0.025392670184373856, + 0.014130868948996067, + 0.02208801358938217, + 0.027378426864743233, + 0.03181681036949158, + 0.021730706095695496, + 0.017730480059981346, + 0.004241269081830978, + 0.00419170456007123, + 0.002911959309130907, + 0.004859553184360266, + 0.006287210620939732, + 0.010594219900667667, + 0.0036789244040846825, + 0.0055772908963263035, + 0.0018005481688305736, + 0.0031541557982563972, + 0.002333960263058543, + 0.003101628739386797, + 0.003209148533642292, + 0.0030066603794693947, + 0.0034501594491302967, + 0.0030820765532553196, + 0.0032368842512369156, + 0.002766040852293372, + 0.0019154658075422049, + 0.002876054495573044, + 0.004147337283939123, + 0.0038837855681777, + 0.0031823869794607162, + 0.003366480115801096, + 0.004474235698580742, + 0.002966377418488264, + 0.0026841359212994576, + 0.005451091565191746, + 0.0035706888884305954, + 0.008185025304555893, + 0.004323351662606001, + 0.0049833557568490505, + 0.007138187997043133, + 0.004684083629399538, + 0.004629151895642281, + 0.00785844773054123, + 0.005186915397644043, + 0.021477947011590004, + 0.005997720640152693, + 0.016873830929398537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.1428101658821106 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979230508208275 + ], + "opposite_logprob_tokens": [ + -3.529792308807373 + ], + "logprob_reward": -0.02979230508208275, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979230508208275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10373599082231522, + 0.048335716128349304, + 0.05083911120891571, + 0.35639944672584534, + 0.12633612751960754, + 0.04849570244550705, + 0.027758419513702393, + 0.03002830222249031, + 0.1341881901025772, + 0.07388301193714142 + ], + "credit_map": [ + 0.010054569691419601, + 0.011263104155659676, + 0.009574918076395988, + 0.01243650447577238, + 0.016826706007122993, + 0.013348724693059921, + 0.0172258373349905, + 0.013005631044507027, + 0.00506052328273654, + 0.00743281003087759, + 0.005125454161316156, + 0.006940627004951239, + 0.005741027183830738, + 0.007654068525880575, + 0.004589206539094448, + 0.0057919989340007305, + 0.005141338799148798, + 0.006686392240226269, + 0.004509086720645428, + 0.005449648946523666, + 0.006627785507589579, + 0.008286912925541401, + 0.008508003316819668, + 0.005629940889775753, + 0.017319872975349426, + 0.033425476402044296, + 0.008829671889543533, + 0.012258145026862621, + 0.036187395453453064, + 0.14255009591579437, + 0.01390570867806673, + 0.0919230654835701, + 0.009261918254196644, + 0.0168254803866148, + 0.007559986785054207, + 0.016468850895762444, + 0.022597866132855415, + 0.02298191748559475, + 0.017138954252004623, + 0.013501142151653767, + 0.004298247862607241, + 0.006359068676829338, + 0.004664601758122444, + 0.005249924026429653, + 0.008935660123825073, + 0.008922800421714783, + 0.004835241939872503, + 0.005230154376477003, + 0.0021818061359226704, + 0.00363367423415184, + 0.0023022887762635946, + 0.002891048789024353, + 0.004142092075198889, + 0.00479489378631115, + 0.004325450398027897, + 0.003487164620310068, + 0.003780944971367717, + 0.005106935277581215, + 0.0034260584507137537, + 0.003841695375740528, + 0.003413422964513302, + 0.0037016281858086586, + 0.003775344928726554, + 0.00298226997256279, + 0.012986443936824799, + 0.011434096843004227, + 0.012791947461664677, + 0.020107006654143333, + 0.018677271902561188, + 0.018367037177085876, + 0.02354465238749981, + 0.01627971976995468, + 0.006775788962841034, + 0.005321095697581768, + 0.003853284055367112, + 0.008033047430217266, + 0.006377892568707466, + 0.021877847611904144, + 0.00578306196257472, + 0.015860991552472115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 4 + ], + "selected_reward": -0.02979230508208275 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic, slow-panning wide shot reveals an otherworldly underwater ballroom, immersed in crystal-clear azure waters. Towering white marble columns, adorned with ancient Greek-style carvings, stand majestically. Between them, almost ghostly, transparent jellyfish pulsate and float gracefully. Their long, translucent tentacles, adorned with bioluminescent algae that emit a soft blue-green glow, serve as an ever-shifting source of light in this serene world. Amidst these luminous creatures, formally dressed individuals—men in classic black tuxedos and women in opulent gowns in deep jewel tones like sapphire and emerald—slowly dance the waltz in near-weightlessness. They twirl and glide gracefully in a dreamy, slow rhythm, their silk and chiffon gowns unfurling and swaying like blossoming flowers in the gentle currents of the water. Above the dance floor, several giant sea turtles glided majestically and silently beneath the ballroom dome, their dark silhouettes casting shimmering, distorted reflections on the highly polished coral floor below, complementing the glow of the dancers and jellyfish to create a tranquil, surreal and magically romantic atmosphere.", + "seed": 26028, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_3/index_8935_seed_26028_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12206767611205578, + "vlm_logprob": -0.12206767611205578, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.12206767611205578, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3525938093662262 + ], + "opposite_logprob_tokens": [ + -1.2275937795639038 + ], + "logprob_reward": -0.3525938093662262, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3525938093662262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08050846308469772, + 0.029257435351610184, + 0.09382698684930801, + 0.18180112540721893, + 0.049909915775060654, + 0.06664375215768814, + 0.03227728605270386, + 0.3136635422706604, + 0.1187886893749237, + 0.0333227664232254 + ], + "credit_map": [ + 0.014500349760055542, + 0.007846761494874954, + 0.011766129173338413, + 0.006623899098485708, + 0.015562754124403, + 0.01268047746270895, + 0.005208215210586786, + 0.006319881416857243, + 0.0043388428166508675, + 0.0030349057633429766, + 0.0026630000211298466, + 0.004780077841132879, + 0.0033893988002091646, + 0.003233903320506215, + 0.00345601630397141, + 0.004361290950328112, + 0.024900246411561966, + 0.010406780987977982, + 0.015223546884953976, + 0.007426472380757332, + 0.01207113265991211, + 0.00881256628781557, + 0.006058456376194954, + 0.008927791379392147, + 0.018452834337949753, + 0.009859481826424599, + 0.04916972666978836, + 0.011002346873283386, + 0.04075020179152489, + 0.020733743906021118, + 0.0100224195048213, + 0.021810369566082954, + 0.004315624013543129, + 0.007176383398473263, + 0.006390413735061884, + 0.005532704293727875, + 0.011374012567102909, + 0.006317480467259884, + 0.004491839557886124, + 0.004311458673328161, + 0.003440629690885544, + 0.01858484372496605, + 0.002907845191657543, + 0.0030826928559690714, + 0.025583216920495033, + 0.005545380990952253, + 0.0040988014079630375, + 0.003400340210646391, + 0.004400069359689951, + 0.004562592599540949, + 0.004272117279469967, + 0.004465226549655199, + 0.004342257976531982, + 0.004588018171489239, + 0.0033662167843431234, + 0.002280788030475378, + 0.02105698548257351, + 0.10334435850381851, + 0.008267264813184738, + 0.019221628084778786, + 0.021420640870928764, + 0.10154043138027191, + 0.02405666932463646, + 0.01475556567311287, + 0.006286063231527805, + 0.035522885620594025, + 0.0033181633334606886, + 0.006539547815918922, + 0.006962108891457319, + 0.012327561154961586, + 0.01005979347974062, + 0.037772566080093384, + 0.0023344815708696842, + 0.005101771093904972, + 0.002476579276844859, + 0.0030483307782560587, + 0.0037166099064052105, + 0.008650842122733593, + 0.00269302143715322, + 0.005301128141582012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.3525938093662262 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020614420995116234 + ], + "opposite_logprob_tokens": [ + -3.8956143856048584 + ], + "logprob_reward": -0.020614420995116234, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020614420995116234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04690015688538551, + 0.04573100060224533, + 0.10296660661697388, + 0.033098459243774414, + 0.030349448323249817, + 0.15678530931472778, + 0.0463617779314518, + 0.3645956814289093, + 0.09600808471441269, + 0.07720346748828888 + ], + "credit_map": [ + 0.008175448514521122, + 0.005373117048293352, + 0.006214885041117668, + 0.0051366956904530525, + 0.006708292756229639, + 0.006086666602641344, + 0.0037950321566313505, + 0.005410023033618927, + 0.007907959632575512, + 0.004963848739862442, + 0.004048334900289774, + 0.006850480102002621, + 0.005693237762898207, + 0.005134613253176212, + 0.005159614607691765, + 0.005972911138087511, + 0.013433140702545643, + 0.011481587775051594, + 0.007326458115130663, + 0.00880648847669363, + 0.014243833720684052, + 0.027341963723301888, + 0.009463803842663765, + 0.010869324207305908, + 0.0036025219596922398, + 0.003015301190316677, + 0.006531492341309786, + 0.002886294387280941, + 0.0072164880111813545, + 0.0036731683649122715, + 0.002668433589860797, + 0.0035047587007284164, + 0.0027921206783503294, + 0.003921513445675373, + 0.0036561358720064163, + 0.004769432824105024, + 0.005185876972973347, + 0.004012081306427717, + 0.0028197872452437878, + 0.003192498814314604, + 0.00830551702529192, + 0.04723219573497772, + 0.004378791432827711, + 0.005104829557240009, + 0.064781554043293, + 0.01290244422852993, + 0.009163129143416882, + 0.00491683091968298, + 0.006811860948801041, + 0.005355530418455601, + 0.004910745657980442, + 0.0046595558524131775, + 0.007327278610318899, + 0.007167476695030928, + 0.005649893078953028, + 0.004479439929127693, + 0.02708381786942482, + 0.11759969592094421, + 0.009928105399012566, + 0.022243551909923553, + 0.026190584525465965, + 0.09930624067783356, + 0.04448498785495758, + 0.017758693546056747, + 0.0049658119678497314, + 0.029570197686553, + 0.0025749895721673965, + 0.004856271669268608, + 0.006102347280830145, + 0.00989140197634697, + 0.008149010129272938, + 0.029898062348365784, + 0.004229054320603609, + 0.010041343979537487, + 0.004610367119312286, + 0.004967426881194115, + 0.006693902425467968, + 0.024944787845015526, + 0.006948766298592091, + 0.014767814427614212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.020614420995116234 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370178863406181 + ], + "opposite_logprob_tokens": [ + -3.4087018966674805 + ], + "logprob_reward": -0.03370178863406181, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370178863406181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04196303337812424, + 0.04155968129634857, + 0.06766689568758011, + 0.10660984367132187, + 0.0906904861330986, + 0.1902829259634018, + 0.01356024295091629, + 0.25507283210754395, + 0.14676612615585327, + 0.0458279624581337 + ], + "credit_map": [ + 0.005602376069873571, + 0.004323374014347792, + 0.007597218733280897, + 0.0047623710706830025, + 0.006749336142092943, + 0.004876138176769018, + 0.003397848689928651, + 0.004654366988688707, + 0.006104474887251854, + 0.0037813615053892136, + 0.004926604218780994, + 0.0065496559254825115, + 0.005232398863881826, + 0.004503070376813412, + 0.004937958903610706, + 0.005524151027202606, + 0.010408778674900532, + 0.006658010184764862, + 0.00497337244451046, + 0.004505622200667858, + 0.010549337603151798, + 0.01958521455526352, + 0.004883496556431055, + 0.006103055085986853, + 0.011499953456223011, + 0.010356592014431953, + 0.023650115355849266, + 0.006722923833876848, + 0.024724507704377174, + 0.011324289254844189, + 0.0055016521364450455, + 0.012829801999032497, + 0.007834217511117458, + 0.01235356554389, + 0.0104090366512537, + 0.016244664788246155, + 0.01401748601347208, + 0.011954229325056076, + 0.008073771372437477, + 0.00980350747704506, + 0.009748765267431736, + 0.05563686043024063, + 0.004727334715425968, + 0.006277109030634165, + 0.07974492758512497, + 0.01638665236532688, + 0.01082449872046709, + 0.006936758756637573, + 0.0016000294126570225, + 0.001978085143491626, + 0.0019138528732582927, + 0.0012607069220393896, + 0.001851651817560196, + 0.002480891766026616, + 0.0014594070380553603, + 0.0010156164644286036, + 0.017183663323521614, + 0.0842113345861435, + 0.0062662833370268345, + 0.015612001530826092, + 0.017476551234722137, + 0.08141975849866867, + 0.020804625004529953, + 0.012098594568669796, + 0.0076019130647182465, + 0.044390447437763214, + 0.003940658178180456, + 0.007858270779252052, + 0.00878021027892828, + 0.015279890038073063, + 0.012465901672840118, + 0.046448808163404465, + 0.0024393696803599596, + 0.005278765223920345, + 0.00227524945512414, + 0.003330579260364175, + 0.0038485387340188026, + 0.01574496179819107, + 0.0026243189349770546, + 0.010286180302500725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.03370178863406181 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603759229183197 + ], + "opposite_logprob_tokens": [ + -1.910375952720642 + ], + "logprob_reward": -0.1603759229183197, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603759229183197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04125193879008293, + 0.02088269777595997, + 0.06461072713136673, + 0.07064376771450043, + 0.05610734969377518, + 0.3514867126941681, + 0.017552092671394348, + 0.3204643726348877, + 0.04394770786166191, + 0.013052570633590221 + ], + "credit_map": [ + 0.005771653261035681, + 0.004611630458384752, + 0.006058027967810631, + 0.004840345121920109, + 0.006126866675913334, + 0.005558851640671492, + 0.002965766005218029, + 0.005318795796483755, + 0.004243829287588596, + 0.001858214265666902, + 0.002583474386483431, + 0.0028416768182069063, + 0.002550952835008502, + 0.002310568932443857, + 0.002100696787238121, + 0.0023932848125696182, + 0.009869038127362728, + 0.006993550341576338, + 0.0045326040126383305, + 0.005069250240921974, + 0.00943820085376501, + 0.018008239567279816, + 0.004885993432253599, + 0.005813847295939922, + 0.00694184610620141, + 0.006525049451738596, + 0.007577484007924795, + 0.006105157546699047, + 0.009723832830786705, + 0.011868453584611416, + 0.006443885155022144, + 0.015458056703209877, + 0.004598288796842098, + 0.007965675555169582, + 0.00671068811789155, + 0.008567246608436108, + 0.009243636392056942, + 0.007938552647829056, + 0.004704138729721308, + 0.006379125639796257, + 0.018027478829026222, + 0.10110308974981308, + 0.008675788529217243, + 0.01166744064539671, + 0.14739353954792023, + 0.02986537478864193, + 0.020175693556666374, + 0.01457833219319582, + 0.0021849502809345722, + 0.0023430201690644026, + 0.0022900858893990517, + 0.0021942558232694864, + 0.002075351309031248, + 0.0030920233111828566, + 0.0018437962280586362, + 0.0015286076813936234, + 0.021522939205169678, + 0.10494130104780197, + 0.007751263678073883, + 0.019805822521448135, + 0.02198549546301365, + 0.10322386771440506, + 0.025214556604623795, + 0.01601913943886757, + 0.0023293427657335997, + 0.012906985357403755, + 0.0013045745436102152, + 0.002528924960643053, + 0.0026664261240512133, + 0.0046356599777936935, + 0.0036971857771277428, + 0.013878608122467995, + 0.0009142368799075484, + 0.002267736941576004, + 0.0013217383529990911, + 0.001697964733466506, + 0.0018078726716339588, + 0.0019997162744402885, + 0.0013874138239771128, + 0.0016558902570977807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.1603759229183197 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an underwater ballroom with towering white marble columns featuring ancient Greek-style carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are formally dressed individuals (men in black tuxedos and women in deep jewel-toned gowns) visible dancing the waltz?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are transparent jellyfish with long translucent tentacles emitting a soft blue-green bioluminescent glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are giant sea turtles gliding silently beneath the dome, casting shimmering reflections on the highly polished coral floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow-panning wide shot revealing the underwater environment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305243864655495 + ], + "opposite_logprob_tokens": [ + -3.1680524349212646 + ], + "logprob_reward": -0.04305243864655495, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305243864655495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022135335952043533, + 0.015317203477025032, + 0.029146796092391014, + 0.08015569299459457, + 0.04225229471921921, + 0.300246924161911, + 0.07812543213367462, + 0.34971922636032104, + 0.06908800452947617, + 0.013813130557537079 + ], + "credit_map": [ + 0.004857951775193214, + 0.00252783065661788, + 0.0024097671266645193, + 0.0024479692801833153, + 0.0032805532682687044, + 0.0028878047596663237, + 0.0016572560416534543, + 0.0020662012975662947, + 0.0033080352004617453, + 0.0024318809155374765, + 0.0012271334417164326, + 0.001683554146438837, + 0.0022157414350658655, + 0.0015937534626573324, + 0.0012272743042558432, + 0.0016298299888148904, + 0.004001629538834095, + 0.0031580468639731407, + 0.0027549949008971453, + 0.0022539144847542048, + 0.004288198426365852, + 0.006667077075690031, + 0.002609390765428543, + 0.0034135424066334963, + 0.005958182737231255, + 0.004534776788204908, + 0.018516357988119125, + 0.005090256221592426, + 0.018114516511559486, + 0.010230808518826962, + 0.003979518078267574, + 0.013731268234550953, + 0.0036519530694931746, + 0.005592457950115204, + 0.0040809400379657745, + 0.006886393763124943, + 0.004566275514662266, + 0.0073456838726997375, + 0.003726613475009799, + 0.0064019798301160336, + 0.01558326929807663, + 0.08921140432357788, + 0.007960989139974117, + 0.009807886555790901, + 0.1258053034543991, + 0.024622062221169472, + 0.017434677109122276, + 0.00982134323567152, + 0.007629545871168375, + 0.01042848639190197, + 0.010686781257390976, + 0.008860195986926556, + 0.012963681481778622, + 0.012355325743556023, + 0.007839459925889969, + 0.007361950818449259, + 0.022106952965259552, + 0.11598905175924301, + 0.008575549349188805, + 0.021394401788711548, + 0.023795928806066513, + 0.12139181792736053, + 0.02012467198073864, + 0.0163408312946558, + 0.0035037221387028694, + 0.02108120545744896, + 0.001891900785267353, + 0.003536237170919776, + 0.004193488974124193, + 0.007143055088818073, + 0.0058486564084887505, + 0.021889738738536835, + 0.0013477494940161705, + 0.002130582695826888, + 0.00139974697958678, + 0.0017412446904927492, + 0.0014151157811284065, + 0.0021414102520793676, + 0.0019339247373864055, + 0.0017033572075888515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.04305243864655495 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_4/index_tensor([2386, 2386])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_4/index_tensor([2386, 2386])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..51be2d3a281438a8fe5d8227b99fc7dbb6f17edf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_4/index_tensor([2386, 2386])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, sideways tracking shot captures the fluidity of Dunhuang murals, depicting an ancient carpet depicting mythical beasts, steadily and calmly navigating the air. Made of thick, handwoven wool, the carpet's aged texture is evident in its slightly frayed edges and crimson tassels that flutter backward in the wind. The carpet's primary colors are aged ochre and lapis lazuli blue, meticulously woven with bold yet fluid calligraphic lines to create an intricate pattern: a graceful, ornately winged nine-colored deer at its center, surrounded by ethereal flying fairies and stylized auspicious clouds, all shimmering with the characteristic matte texture of mineral pigments. The carpet, rather than moving in a rigid, lifelike motion, rises and falls steadily through the air, as if sailing on invisible waves. Below the camera, a vast expanse of golden desert stretches to the horizon. The setting sun paints the rolling dunes a warm, flowing sea of gold, casting long, soft blue-purple shadows on the leeward sides. The sky, far from a realistic depiction, is a splendid azure, punctuated by elongated, brushstroke-like orange-red clouds. The entire scene is bathed in warm and soft light, imbued with a sacred, tranquil, and epic atmosphere, as if a millennia-old mural has been brought to life and is slowly unfolding before our eyes.", + "seed": 6349, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_4/index_2386_seed_6349_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.033362907357513905, + "vlm_logprob": -0.033362907357513905, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.033362907357513905, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024606289342045784 + ], + "opposite_logprob_tokens": [ + -3.774606227874756 + ], + "logprob_reward": -0.024606289342045784, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.024606289342045784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09601497650146484, + 0.07479635626077652, + 0.0793289765715599, + 0.05091092362999916, + 0.1486029475927353, + 0.0453103706240654, + 0.09442251175642014, + 0.17364943027496338, + 0.16610078513622284, + 0.07086263597011566 + ], + "credit_map": [ + 0.010095587000250816, + 0.008567449636757374, + 0.014582422561943531, + 0.017041662707924843, + 0.007786743342876434, + 0.01320840697735548, + 0.010130388662219048, + 0.014602315612137318, + 0.008213216438889503, + 0.008212652988731861, + 0.009039643220603466, + 0.007532711606472731, + 0.007761318702250719, + 0.013704270124435425, + 0.010360859334468842, + 0.00997168943285942, + 0.006917321588844061, + 0.008431673049926758, + 0.01222500391304493, + 0.009749012067914009, + 0.013202302157878876, + 0.006546028424054384, + 0.011550799012184143, + 0.010706842876970768, + 0.004792746156454086, + 0.006575047504156828, + 0.005390456411987543, + 0.009413032792508602, + 0.005808133631944656, + 0.0048547424376010895, + 0.006492678541690111, + 0.00758409034460783, + 0.015178719535470009, + 0.020873134955763817, + 0.021869491785764694, + 0.01408557128161192, + 0.021537523716688156, + 0.0088026849552989, + 0.013472574763000011, + 0.03278326988220215, + 0.004649120382964611, + 0.006057959981262684, + 0.006924312096089125, + 0.007043934892863035, + 0.006562709342688322, + 0.004089062102138996, + 0.004691367503255606, + 0.005291908513754606, + 0.006499612703919411, + 0.011491846293210983, + 0.01157464925199747, + 0.012742684222757816, + 0.016667669638991356, + 0.008012077771127224, + 0.010444610379636288, + 0.01698937825858593, + 0.0342431515455246, + 0.013520154170691967, + 0.06346473842859268, + 0.015517243184149265, + 0.019118957221508026, + 0.007857554592192173, + 0.008942665532231331, + 0.010984988883137703, + 0.028993388637900352, + 0.010626624338328838, + 0.058937955647706985, + 0.013108786195516586, + 0.018302978947758675, + 0.00816186610609293, + 0.010781124234199524, + 0.017188098281621933, + 0.007446556352078915, + 0.007438899017870426, + 0.008109214715659618, + 0.012204675003886223, + 0.010171864181756973, + 0.007169506978243589, + 0.0074839284643530846, + 0.010837997309863567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.024606289342045784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607024483382702 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.01607024483382702, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607024483382702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1808580756187439, + 0.19750572741031647, + 0.0844230055809021, + 0.07393605262041092, + 0.1383981704711914, + 0.07510877400636673, + 0.03050677850842476, + 0.13352519273757935, + 0.047664910554885864, + 0.03807326778769493 + ], + "credit_map": [ + 0.018256554380059242, + 0.020549140870571136, + 0.030156536027789116, + 0.026381589472293854, + 0.01675640232861042, + 0.02100350707769394, + 0.020868046209216118, + 0.02688630297780037, + 0.02116885408759117, + 0.027897648513317108, + 0.029841415584087372, + 0.025122210383415222, + 0.021602841094136238, + 0.025160130113363266, + 0.02231317013502121, + 0.02439946122467518, + 0.00798091758042574, + 0.01029966026544571, + 0.013050700537860394, + 0.009064258076250553, + 0.018091654404997826, + 0.006385112646967173, + 0.01192663423717022, + 0.007624064572155476, + 0.006945135071873665, + 0.006882075686007738, + 0.011304216459393501, + 0.020916858687996864, + 0.007564496248960495, + 0.005606474354863167, + 0.006708221510052681, + 0.008008576929569244, + 0.01565183512866497, + 0.009055371396243572, + 0.018700746819376945, + 0.010196637362241745, + 0.012951490469276905, + 0.007631191983819008, + 0.011661091819405556, + 0.0525498129427433, + 0.00899040699005127, + 0.010174525901675224, + 0.014350509271025658, + 0.00760706327855587, + 0.01211556140333414, + 0.005556571763008833, + 0.007163868751376867, + 0.009150263853371143, + 0.0028850582893937826, + 0.0037981709465384483, + 0.004408986773341894, + 0.003536453004926443, + 0.005086031276732683, + 0.0030094319954514503, + 0.003334204899147153, + 0.004448441788554192, + 0.01554061844944954, + 0.020838208496570587, + 0.024460094049572945, + 0.016796495765447617, + 0.01943778246641159, + 0.014158036559820175, + 0.010350247845053673, + 0.011943701654672623, + 0.008125451393425465, + 0.004011158365756273, + 0.015032622963190079, + 0.003925500903278589, + 0.005034925881773233, + 0.0028066285885870457, + 0.0030670862179249525, + 0.0056615364737808704, + 0.005263195373117924, + 0.004207534249871969, + 0.0073087275959551334, + 0.0059466916136443615, + 0.004776632413268089, + 0.002520753303542733, + 0.0036890290211886168, + 0.004360707942396402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.01607024483382702 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606355793774128 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.01606355793774128, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606355793774128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16746196150779724, + 0.06485239416360855, + 0.07917537540197372, + 0.04610291123390198, + 0.3029826879501343, + 0.03502567857503891, + 0.0420742966234684, + 0.16816040873527527, + 0.05502187833189964, + 0.039142463356256485 + ], + "credit_map": [ + 0.01750047691166401, + 0.023580867797136307, + 0.016629748046398163, + 0.020282911136746407, + 0.02747483178973198, + 0.02531306818127632, + 0.019727496430277824, + 0.016952570527791977, + 0.00677521713078022, + 0.00780908390879631, + 0.01060988288372755, + 0.007660302799195051, + 0.006605606060475111, + 0.009440571069717407, + 0.00752171128988266, + 0.008430018089711666, + 0.008485308848321438, + 0.008499209769070148, + 0.012192053720355034, + 0.009629196487367153, + 0.013956679962575436, + 0.00581733463332057, + 0.009528013877570629, + 0.01106757577508688, + 0.00449135759845376, + 0.006006165407598019, + 0.0064047533087432384, + 0.00983806699514389, + 0.006937344558537006, + 0.004025445319712162, + 0.003620434319600463, + 0.00477934442460537, + 0.034082889556884766, + 0.0226645078510046, + 0.048627935349941254, + 0.01753305085003376, + 0.02637283317744732, + 0.016955096274614334, + 0.021472301334142685, + 0.11527407169342041, + 0.002752144355326891, + 0.004171514883637428, + 0.00671053072437644, + 0.004714976064860821, + 0.004128854721784592, + 0.0033151432871818542, + 0.003500107442960143, + 0.0057324073277413845, + 0.003147390205413103, + 0.005280546378344297, + 0.005492825992405415, + 0.004685075022280216, + 0.008427278138697147, + 0.004112841561436653, + 0.004420176614075899, + 0.00650816410779953, + 0.03357957303524017, + 0.011743850074708462, + 0.06350257992744446, + 0.014317815192043781, + 0.017074251547455788, + 0.007466485258191824, + 0.008397392928600311, + 0.012078470550477505, + 0.007866065949201584, + 0.00637913728132844, + 0.011290756054222584, + 0.0053355093114078045, + 0.006759065669029951, + 0.004437657538801432, + 0.004944671876728535, + 0.008009010925889015, + 0.0061231148429214954, + 0.0031980325002223253, + 0.00873416569083929, + 0.005090699996799231, + 0.005604938138276339, + 0.0027796123176813126, + 0.003217643592506647, + 0.00439425790682435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.01606355793774128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030708461999893 + ], + "opposite_logprob_tokens": [ + -2.350306987762451 + ], + "logprob_reward": -0.10030708461999893, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030708461999893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18667255342006683, + 0.06066572293639183, + 0.09936441481113434, + 0.03337148576974869, + 0.37032997608184814, + 0.044434115290641785, + 0.02666686661541462, + 0.04468847066164017, + 0.08643096685409546, + 0.047375354915857315 + ], + "credit_map": [ + 0.017108701169490814, + 0.021360602229833603, + 0.026957977563142776, + 0.02335307188332081, + 0.02012939751148224, + 0.03017144650220871, + 0.020192889496684074, + 0.027398476377129555, + 0.006326272618025541, + 0.006209293846040964, + 0.009337915107607841, + 0.00805442500859499, + 0.005745184142142534, + 0.011688265018165112, + 0.007452972698956728, + 0.00585140148177743, + 0.008537006564438343, + 0.009641137905418873, + 0.01307876780629158, + 0.013320562429726124, + 0.018474703654646873, + 0.010151668451726437, + 0.015232378616929054, + 0.010928197763860226, + 0.0031780023127794266, + 0.0033859524410218, + 0.005674393847584724, + 0.006881350185722113, + 0.003805232234299183, + 0.003271416062489152, + 0.002761612879112363, + 0.004413527436554432, + 0.04682764783501625, + 0.02344362437725067, + 0.05331791564822197, + 0.021095048636198044, + 0.029472336173057556, + 0.0218326598405838, + 0.02723383530974388, + 0.14710694551467896, + 0.0034414073452353477, + 0.006148879881948233, + 0.004836174193769693, + 0.006158617325127125, + 0.008720331825315952, + 0.0061895232647657394, + 0.004409819841384888, + 0.0045293644070625305, + 0.002214707899838686, + 0.0024987102951854467, + 0.004516714718192816, + 0.0034254633355885744, + 0.006184480153024197, + 0.0023172590881586075, + 0.0029350793920457363, + 0.002574453130364418, + 0.00717704463750124, + 0.0037017895374447107, + 0.013939710333943367, + 0.0037562535144388676, + 0.005419107154011726, + 0.003390320809558034, + 0.00324118766002357, + 0.004063059110194445, + 0.015743272379040718, + 0.006504385266453028, + 0.030479298904538155, + 0.005602690391242504, + 0.009775988757610321, + 0.0037922330666333437, + 0.005686930846422911, + 0.008846175856888294, + 0.0055951206013560295, + 0.004604386631399393, + 0.009372247382998466, + 0.006181242875754833, + 0.009461061097681522, + 0.004597478546202183, + 0.00395843293517828, + 0.003605388570576906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.10030708461999893 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009767360053956509 + ], + "opposite_logprob_tokens": [ + -4.634767532348633 + ], + "logprob_reward": -0.009767360053956509, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009767360053956509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1409069448709488, + 0.023627925664186478, + 0.06200837716460228, + 0.06581716239452362, + 0.05596677213907242, + 0.05840516462922096, + 0.04839753359556198, + 0.43077489733695984, + 0.046066250652074814, + 0.0680289939045906 + ], + "credit_map": [ + 0.013634593226015568, + 0.012756110168993473, + 0.020527144894003868, + 0.014706783927977085, + 0.01606186106801033, + 0.02591830864548683, + 0.014754898846149445, + 0.022547248750925064, + 0.0023198919370770454, + 0.0029832657892256975, + 0.0036413550842553377, + 0.0021506918128579855, + 0.0030896179378032684, + 0.0031780321151018143, + 0.002720208838582039, + 0.0035448619164526463, + 0.005832897964864969, + 0.00498995603993535, + 0.006677713710814714, + 0.009726333431899548, + 0.010715059004724026, + 0.006315500941127539, + 0.012100466527044773, + 0.005650448612868786, + 0.006310776807367802, + 0.005713476333767176, + 0.008209793828427792, + 0.01826462335884571, + 0.006160541903227568, + 0.005646486766636372, + 0.0081546101719141, + 0.007356856018304825, + 0.004196920432150364, + 0.005071492865681648, + 0.011585529893636703, + 0.005085573066025972, + 0.0066262902691960335, + 0.004482485353946686, + 0.004581383429467678, + 0.01433709729462862, + 0.006054861005395651, + 0.007109908852726221, + 0.008897917345166206, + 0.008872613310813904, + 0.007412997540086508, + 0.005648490507155657, + 0.006227354984730482, + 0.008181020617485046, + 0.004397301003336906, + 0.008101641200482845, + 0.00826176069676876, + 0.005200717132538557, + 0.007881184108555317, + 0.004180570133030415, + 0.005029794294387102, + 0.005344563629478216, + 0.09113378077745438, + 0.028815215453505516, + 0.1721571385860443, + 0.03112725540995598, + 0.04224315658211708, + 0.017414515838027, + 0.016758140176534653, + 0.031125713139772415, + 0.008823993615806103, + 0.003647386096417904, + 0.01761777140200138, + 0.0036682577338069677, + 0.0046623749658465385, + 0.0016096982872113585, + 0.002310494426637888, + 0.003726277034729719, + 0.01418562326580286, + 0.004905289970338345, + 0.02379898726940155, + 0.005921898875385523, + 0.008293409831821918, + 0.002714213216677308, + 0.0032114400528371334, + 0.004998132586479187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.009767360053956509 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, sideways tracking shot captures the fluidity of Dunhuang murals, depicting an ancient carpet depicting mythical beasts, steadily and calmly navigating the air. Made of thick, handwoven wool, the carpet's aged texture is evident in its slightly frayed edges and crimson tassels that flutter backward in the wind. The carpet's primary colors are aged ochre and lapis lazuli blue, meticulously woven with bold yet fluid calligraphic lines to create an intricate pattern: a graceful, ornately winged nine-colored deer at its center, surrounded by ethereal flying fairies and stylized auspicious clouds, all shimmering with the characteristic matte texture of mineral pigments. The carpet, rather than moving in a rigid, lifelike motion, rises and falls steadily through the air, as if sailing on invisible waves. Below the camera, a vast expanse of golden desert stretches to the horizon. The setting sun paints the rolling dunes a warm, flowing sea of gold, casting long, soft blue-purple shadows on the leeward sides. The sky, far from a realistic depiction, is a splendid azure, punctuated by elongated, brushstroke-like orange-red clouds. The entire scene is bathed in warm and soft light, imbued with a sacred, tranquil, and epic atmosphere, as if a millennia-old mural has been brought to life and is slowly unfolding before our eyes.", + "seed": 6350, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_4/index_2386_seed_6350_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03285715992096812, + "vlm_logprob": -0.03285715992096812, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03285715992096812, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09049078822135925 + ], + "opposite_logprob_tokens": [ + -2.4654908180236816 + ], + "logprob_reward": -0.09049078822135925, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09049078822135925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.943364679813385, + 0.009721214883029461, + 0.009125081822276115, + 0.008630256168544292, + 0.003268411150202155, + 0.0020138979889452457, + 0.010521731339395046, + 0.0016198311932384968, + 0.009221975691616535, + 0.002513033105060458 + ], + "credit_map": [ + 0.09931397438049316, + 0.03285655379295349, + 0.3358558416366577, + 0.06559336185455322, + 0.02720077522099018, + 0.22300182282924652, + 0.05222247913479805, + 0.10731983929872513, + 0.0014967566821724176, + 0.0009388174512423575, + 0.0011875844793394208, + 0.0014560751151293516, + 0.0007957975030876696, + 0.000910333706997335, + 0.0019185919081792235, + 0.0010172585025429726, + 0.0009190784767270088, + 0.0009246896370314062, + 0.0008447744767181575, + 0.0021191355772316456, + 0.0010272503132000566, + 0.0009484725305810571, + 0.001732606440782547, + 0.0006090747774578631, + 0.0007550856680609286, + 0.0009777159430086613, + 0.0007717139669694006, + 0.0016077934997156262, + 0.0016043903306126595, + 0.0006304191192612052, + 0.0014869202859699726, + 0.0007962182280607522, + 0.0002693743153940886, + 0.0004922769730910659, + 0.0002537586260586977, + 0.0006606393144465983, + 0.000718951050657779, + 0.0002623397740535438, + 0.0003442652814555913, + 0.0002668056695256382, + 0.0001918916095746681, + 0.00024010345805436373, + 0.0002892968477681279, + 0.00028216911596246064, + 0.00031042160117067397, + 0.00015058330609463155, + 0.00026151741622015834, + 0.00028791482327505946, + 0.0009718743385747075, + 0.0015225165989249945, + 0.0010902935173362494, + 0.001354797976091504, + 0.0023776942398399115, + 0.0009780755499377847, + 0.0009777683299034834, + 0.0012487109052017331, + 0.0001610993786016479, + 0.0002589036012068391, + 0.00018127112707588822, + 0.0002664149214979261, + 0.00024093348474707454, + 0.0001262177393073216, + 0.00020057836081832647, + 0.00018441250722389668, + 0.0010735398391261697, + 0.0009426164906471968, + 0.0009955394780263305, + 0.0008517930982634425, + 0.0022375064436346292, + 0.0010782865574583411, + 0.0009319286327809095, + 0.0011107648024335504, + 0.00030365120619535446, + 0.0003185678506270051, + 0.00031260133255273104, + 0.00033311120932921767, + 0.00031204448896460235, + 0.0001768118527252227, + 0.0003734170750249177, + 0.0003828280314337462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.09049078822135925 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589079707860947 + ], + "opposite_logprob_tokens": [ + -3.8955891132354736 + ], + "logprob_reward": -0.020589079707860947, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589079707860947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.819328784942627, + 0.026580220088362694, + 0.02234623022377491, + 0.03685225546360016, + 0.015946989879012108, + 0.01475352980196476, + 0.030643370002508163, + 0.009427943266928196, + 0.010337634943425655, + 0.013782999478280544 + ], + "credit_map": [ + 0.0845622792840004, + 0.02889031358063221, + 0.29447469115257263, + 0.05723325535655022, + 0.023882150650024414, + 0.1909332275390625, + 0.04617464914917946, + 0.09317824989557266, + 0.00297976890578866, + 0.00431471224874258, + 0.0034106371458619833, + 0.004244672134518623, + 0.0026582933496683836, + 0.0024278860073536634, + 0.0038022948428988457, + 0.0027419556863605976, + 0.001742420601658523, + 0.002109555061906576, + 0.0028510543052107096, + 0.005518911872059107, + 0.0020850938744843006, + 0.0016115957405418158, + 0.004526511766016483, + 0.0019010868854820728, + 0.003266759682446718, + 0.004752230830490589, + 0.004556943662464619, + 0.00893409177660942, + 0.003628153121098876, + 0.003071080893278122, + 0.00526402797549963, + 0.003378968220204115, + 0.001464071567170322, + 0.002912609837949276, + 0.0017179016722366214, + 0.002357128541916609, + 0.003156656865030527, + 0.0011390363797545433, + 0.001977607375010848, + 0.0012219775235280395, + 0.0014522698475047946, + 0.0031943232752382755, + 0.0016378586879000068, + 0.0016927743563428521, + 0.002316372701898217, + 0.0009744228445924819, + 0.0016735660610720515, + 0.0018119419692084193, + 0.0027644529473036528, + 0.004649804439395666, + 0.002995301503688097, + 0.004034293349832296, + 0.006859026849269867, + 0.0028355352114886045, + 0.0029240045696496964, + 0.003580948803573847, + 0.0013592906761914492, + 0.000913616269826889, + 0.0010330949444323778, + 0.001043760566972196, + 0.0019868183881044388, + 0.0008245630306191742, + 0.0008719209581613541, + 0.0013948788400739431, + 0.001083770883269608, + 0.001080645713955164, + 0.0010776419658213854, + 0.0009409843478351831, + 0.0025929308030754328, + 0.0012170978588983417, + 0.0009761161054484546, + 0.0013684481382369995, + 0.0017372866859659553, + 0.0016323718009516597, + 0.0019085590029135346, + 0.0013952016597613692, + 0.00197375169955194, + 0.0013251029886305332, + 0.0017354715382680297, + 0.002075253054499626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.020589079707860947 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636386133730412 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636386133730412, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636386133730412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8637602925300598, + 0.014217115938663483, + 0.012399690225720406, + 0.021678855642676353, + 0.010582064278423786, + 0.010884659364819527, + 0.032852813601493835, + 0.006453004665672779, + 0.016765454784035683, + 0.010405982844531536 + ], + "credit_map": [ + 0.09075190126895905, + 0.030418794602155685, + 0.3049045503139496, + 0.06051122769713402, + 0.025173047557473183, + 0.20455768704414368, + 0.049392104148864746, + 0.09805099666118622, + 0.0018355189822614193, + 0.0015405794838443398, + 0.0016280179843306541, + 0.0021326954010874033, + 0.0013535645557567477, + 0.0014743957435712218, + 0.002541618188843131, + 0.00171072653029114, + 0.001410280354321003, + 0.0015536075225099921, + 0.001955382525920868, + 0.0023740583565086126, + 0.0012064597103744745, + 0.0011148561025038362, + 0.0017702069599181414, + 0.0010148384608328342, + 0.0017203923780471087, + 0.0026720883324742317, + 0.002195845590904355, + 0.0049833995290100574, + 0.0021202347707003355, + 0.001442764769308269, + 0.003729462856426835, + 0.0028146677650511265, + 0.0010331111261621118, + 0.0015917406417429447, + 0.0010289173806086183, + 0.00171139114536345, + 0.002597528975456953, + 0.0008296911837533116, + 0.0011100973933935165, + 0.0006795868393965065, + 0.0011442819377407432, + 0.0016900351038202643, + 0.001486359746195376, + 0.0012666325783357024, + 0.002116918098181486, + 0.0007826727814972401, + 0.0012900468427687883, + 0.0011077129747718573, + 0.0032897545024752617, + 0.0046347277238965034, + 0.0034762287978082895, + 0.004436178132891655, + 0.006957871373742819, + 0.003088582307100296, + 0.003180786734446883, + 0.0037886842619627714, + 0.0009837381076067686, + 0.0005816906341351569, + 0.0007503457018174231, + 0.0006683961837552488, + 0.0015174574218690395, + 0.000583305605687201, + 0.0006655926117673516, + 0.000702478107996285, + 0.0018229661509394646, + 0.0016452837735414505, + 0.0017308401875197887, + 0.0013867600355297327, + 0.004293179605156183, + 0.001995947677642107, + 0.0018190249102190137, + 0.002071452559903264, + 0.0010694029042497277, + 0.0011371001601219177, + 0.0015158815076574683, + 0.0012567553203552961, + 0.0010064617963507771, + 0.0008985314052551985, + 0.0012167055392637849, + 0.0023051437456160784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.008636386133730412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306066036224365 + ], + "opposite_logprob_tokens": [ + -3.168060779571533 + ], + "logprob_reward": -0.04306066036224365, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306066036224365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6300085186958313, + 0.07167346775531769, + 0.03380042687058449, + 0.034720808267593384, + 0.04834548756480217, + 0.0212753564119339, + 0.10236243903636932, + 0.019765475764870644, + 0.01184141356498003, + 0.026206644251942635 + ], + "credit_map": [ + 0.06572999805212021, + 0.02306017093360424, + 0.2072315365076065, + 0.0427892729640007, + 0.028263600543141365, + 0.1382160186767578, + 0.05316924676299095, + 0.07154866307973862, + 0.00731345871463418, + 0.010695474222302437, + 0.009898186661303043, + 0.014379299245774746, + 0.006716008298099041, + 0.006233162246644497, + 0.010343830101191998, + 0.0060940454714000225, + 0.002831305144354701, + 0.0034660627134144306, + 0.003436434082686901, + 0.008093723095953465, + 0.002605422167107463, + 0.0035495911724865437, + 0.007041933946311474, + 0.002775951288640499, + 0.00276047526858747, + 0.004333739168941975, + 0.004029173869639635, + 0.0066847968846559525, + 0.004466366954147816, + 0.0026287161745131016, + 0.005900693126022816, + 0.003916846588253975, + 0.0044883377850055695, + 0.008177884854376316, + 0.005365916993469, + 0.008699274621903896, + 0.0067062415182590485, + 0.0038868950214236975, + 0.005459302570670843, + 0.005561631638556719, + 0.0016825059428811073, + 0.002748149214312434, + 0.003189545590430498, + 0.0031674420461058617, + 0.0029143900610506535, + 0.0015677929623052478, + 0.002449579071253538, + 0.003555949777364731, + 0.009452667087316513, + 0.014867421239614487, + 0.009550998918712139, + 0.013678964227437973, + 0.022949375212192535, + 0.009698200039565563, + 0.009918341413140297, + 0.012246464379131794, + 0.0027562256436794996, + 0.00219447142444551, + 0.0018207578686997294, + 0.002117076888680458, + 0.004569011740386486, + 0.0018232518341392279, + 0.0018805335275828838, + 0.002604142762720585, + 0.0013668289175257087, + 0.0013292914954945445, + 0.001537042553536594, + 0.0015661228680983186, + 0.0022104375530034304, + 0.0012614702573046088, + 0.0012969608651474118, + 0.001273258705623448, + 0.0024666120298206806, + 0.0024287495762109756, + 0.0026717307046055794, + 0.0026212662924081087, + 0.002411325927823782, + 0.001957228872925043, + 0.004354094620794058, + 0.007295633666217327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.04306066036224365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015088851796463132 + ], + "opposite_logprob_tokens": [ + -6.501508712768555 + ], + "logprob_reward": -0.0015088851796463132, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015088851796463132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8648990988731384, + 0.03538239374756813, + 0.0107243862003088, + 0.012983494438230991, + 0.007326945662498474, + 0.014729778282344341, + 0.016034219413995743, + 0.007866565138101578, + 0.006285776849836111, + 0.023767422884702682 + ], + "credit_map": [ + 0.09228423237800598, + 0.029756981879472733, + 0.30624037981033325, + 0.05922580882906914, + 0.025882981717586517, + 0.2056194394826889, + 0.047955725342035294, + 0.09793350100517273, + 0.004133785609155893, + 0.005619061645120382, + 0.005874058231711388, + 0.005614157300442457, + 0.003251845482736826, + 0.0031408709473907948, + 0.004543687216937542, + 0.0032049273140728474, + 0.0015491573140025139, + 0.0013073516311123967, + 0.0013422026531770825, + 0.001716917147859931, + 0.0010112877935171127, + 0.0012838339898735285, + 0.0016004852950572968, + 0.0009131492697633803, + 0.0013358310097828507, + 0.0011811589356511831, + 0.0018158818129450083, + 0.001937606604769826, + 0.00262350938282907, + 0.00109980592969805, + 0.0015653346199542284, + 0.0014243641635403037, + 0.000624552892986685, + 0.0009263842948712409, + 0.0007675777887925506, + 0.001433994504623115, + 0.001706001814454794, + 0.0006294231279753149, + 0.0008058684761635959, + 0.0004331414820626378, + 0.0016118615167215466, + 0.0023071784526109695, + 0.002161909593269229, + 0.001900527160614729, + 0.0018081867601722479, + 0.0009810198098421097, + 0.0018694495083764195, + 0.002089644782245159, + 0.0015088891377672553, + 0.0021690367721021175, + 0.0016809668159112334, + 0.0021434787195175886, + 0.0036306215915828943, + 0.001406448194757104, + 0.00152850232552737, + 0.001966273644939065, + 0.001361239468678832, + 0.0006792982458136976, + 0.000982971745543182, + 0.0008160321740433574, + 0.001511291484348476, + 0.000626563502009958, + 0.0007352831889875233, + 0.0011538846883922815, + 0.0008019796805456281, + 0.0006096958531998098, + 0.0007882940117269754, + 0.0005700847832486033, + 0.0014712498523294926, + 0.0006953650154173374, + 0.0007153795449994504, + 0.0006337272934615612, + 0.0026499880477786064, + 0.0024327035062015057, + 0.0028069913387298584, + 0.0023732148110866547, + 0.003461843356490135, + 0.0023827964905649424, + 0.00378771242685616, + 0.0038721708115190268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0015088851796463132 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, sideways tracking shot captures the fluidity of Dunhuang murals, depicting an ancient carpet depicting mythical beasts, steadily and calmly navigating the air. Made of thick, handwoven wool, the carpet's aged texture is evident in its slightly frayed edges and crimson tassels that flutter backward in the wind. The carpet's primary colors are aged ochre and lapis lazuli blue, meticulously woven with bold yet fluid calligraphic lines to create an intricate pattern: a graceful, ornately winged nine-colored deer at its center, surrounded by ethereal flying fairies and stylized auspicious clouds, all shimmering with the characteristic matte texture of mineral pigments. The carpet, rather than moving in a rigid, lifelike motion, rises and falls steadily through the air, as if sailing on invisible waves. Below the camera, a vast expanse of golden desert stretches to the horizon. The setting sun paints the rolling dunes a warm, flowing sea of gold, casting long, soft blue-purple shadows on the leeward sides. The sky, far from a realistic depiction, is a splendid azure, punctuated by elongated, brushstroke-like orange-red clouds. The entire scene is bathed in warm and soft light, imbued with a sacred, tranquil, and epic atmosphere, as if a millennia-old mural has been brought to life and is slowly unfolding before our eyes.", + "seed": 6351, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_4/index_2386_seed_6351_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028522223606705664, + "vlm_logprob": -0.028522223606705664, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.028522223606705664, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014040032401680946 + ], + "opposite_logprob_tokens": [ + -4.389039993286133 + ], + "logprob_reward": -0.014040032401680946, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014040032401680946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1531781703233719, + 0.06154650077223778, + 0.06153242662549019, + 0.14520664513111115, + 0.07814866304397583, + 0.043253056704998016, + 0.050368230789899826, + 0.10764044523239136, + 0.13510239124298096, + 0.1640234738588333 + ], + "credit_map": [ + 0.016671709716320038, + 0.01978881098330021, + 0.02436099201440811, + 0.020391587167978287, + 0.01753009669482708, + 0.017965395003557205, + 0.015531769953668118, + 0.020937804132699966, + 0.00856594368815422, + 0.01616777293384075, + 0.00628468906506896, + 0.006100145634263754, + 0.0066366055980324745, + 0.006734183989465237, + 0.005977018736302853, + 0.005080142989754677, + 0.00729874474927783, + 0.013411344029009342, + 0.006141895893961191, + 0.005925890523940325, + 0.007614720147103071, + 0.008073038421571255, + 0.006196831353008747, + 0.006869960576295853, + 0.017355799674987793, + 0.0138847129419446, + 0.013818896375596523, + 0.011711019091308117, + 0.022816220298409462, + 0.015660889446735382, + 0.02284950017929077, + 0.02710961364209652, + 0.008031868375837803, + 0.010498974472284317, + 0.010943914763629436, + 0.010297814384102821, + 0.011296169832348824, + 0.00895405188202858, + 0.010234590619802475, + 0.007891277782619, + 0.004713176749646664, + 0.0050847092643380165, + 0.0080260019749403, + 0.003928481601178646, + 0.004732861649245024, + 0.006829485762864351, + 0.005452399607747793, + 0.004485940095037222, + 0.004718664102256298, + 0.006491281557828188, + 0.005831514485180378, + 0.004662518855184317, + 0.00902478490024805, + 0.006179395597428083, + 0.007937618531286716, + 0.005522451363503933, + 0.008508000522851944, + 0.011157812550663948, + 0.00891420803964138, + 0.010188605636358261, + 0.023087723180651665, + 0.01085357740521431, + 0.02497953735291958, + 0.00995098426938057, + 0.01477825827896595, + 0.01792304776608944, + 0.017132561653852463, + 0.011542292311787605, + 0.030951565131545067, + 0.012853005900979042, + 0.01834091916680336, + 0.01158074475824833, + 0.01706075854599476, + 0.015147246420383453, + 0.014565898105502129, + 0.01888689398765564, + 0.015887433663010597, + 0.014285392127931118, + 0.050725359469652176, + 0.017464490607380867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.014040032401680946 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289121687412262 + ], + "opposite_logprob_tokens": [ + -3.773289203643799 + ], + "logprob_reward": -0.023289121687412262, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289121687412262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13723596930503845, + 0.04961448162794113, + 0.029590098187327385, + 0.15288837254047394, + 0.06381293386220932, + 0.017233148217201233, + 0.05873247981071472, + 0.08065996319055557, + 0.08849076181650162, + 0.32174181938171387 + ], + "credit_map": [ + 0.01611749827861786, + 0.030557943508028984, + 0.023617396131157875, + 0.020289503037929535, + 0.014737236313521862, + 0.010310503654181957, + 0.013024486601352692, + 0.00858138594776392, + 0.005330526735633612, + 0.008183618076145649, + 0.00798744522035122, + 0.006629433948546648, + 0.005955271422863007, + 0.004877916071563959, + 0.006180993281304836, + 0.00446927547454834, + 0.0038270256482064724, + 0.005514621268957853, + 0.003829154884442687, + 0.0025787486229091883, + 0.003886784426867962, + 0.0029235936235636473, + 0.004021116998046637, + 0.003009048756211996, + 0.019992506131529808, + 0.015027216635644436, + 0.01295736338943243, + 0.011816933751106262, + 0.023139482364058495, + 0.017306536436080933, + 0.02299484796822071, + 0.029653480276465416, + 0.006276795640587807, + 0.008579278364777565, + 0.010685618035495281, + 0.009032267145812511, + 0.008345484733581543, + 0.007349980529397726, + 0.006913935765624046, + 0.006629565265029669, + 0.001927984762005508, + 0.0023017420899122953, + 0.002601753920316696, + 0.0017803343944251537, + 0.0024034392554312944, + 0.002196836518123746, + 0.0025946791283786297, + 0.0014263769844546914, + 0.005148696713149548, + 0.008984212763607502, + 0.008875717408955097, + 0.0047556618228554726, + 0.007319629658013582, + 0.006722648162394762, + 0.008189988322556019, + 0.008735920302569866, + 0.010302065871655941, + 0.009875847026705742, + 0.008101314306259155, + 0.00612719776108861, + 0.017142195254564285, + 0.010895101353526115, + 0.011878189630806446, + 0.0063380468636751175, + 0.006903705187141895, + 0.009543162770569324, + 0.021331652998924255, + 0.008398530073463917, + 0.010715460404753685, + 0.010528473183512688, + 0.010760780423879623, + 0.010308995842933655, + 0.038902148604393005, + 0.020052777603268623, + 0.02038947306573391, + 0.039277736097574234, + 0.023644890636205673, + 0.014477885328233242, + 0.13158831000328064, + 0.03340859338641167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.023289121687412262 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635440841317177 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635440841317177, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635440841317177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1352747529745102, + 0.06974276155233383, + 0.034351058304309845, + 0.03418104723095894, + 0.028535041958093643, + 0.03100714646279812, + 0.03551945090293884, + 0.3565709888935089, + 0.09527714550495148, + 0.17954064905643463 + ], + "credit_map": [ + 0.01489141583442688, + 0.027123985812067986, + 0.021757079288363457, + 0.01628713496029377, + 0.014360913075506687, + 0.013256958685815334, + 0.01588173769414425, + 0.011715524829924107, + 0.007461200468242168, + 0.010874845087528229, + 0.013264342211186886, + 0.010669901967048645, + 0.006834620609879494, + 0.006320581771433353, + 0.007897468283772469, + 0.006419803015887737, + 0.0039649223908782005, + 0.005084764212369919, + 0.003921580966562033, + 0.0029608907643705606, + 0.004605582915246487, + 0.006470200605690479, + 0.004053595010191202, + 0.0032895186450332403, + 0.003383100498467684, + 0.004860223270952702, + 0.0032144575379788876, + 0.004018050152808428, + 0.005304343067109585, + 0.003935612738132477, + 0.004938790574669838, + 0.004526468925178051, + 0.002779861679300666, + 0.002925726817920804, + 0.004432652145624161, + 0.00301918457262218, + 0.0047554015181958675, + 0.0028689568862318993, + 0.004950571805238724, + 0.002802689792588353, + 0.0038704059552401304, + 0.004069264512509108, + 0.0037173228338360786, + 0.0031506812665611506, + 0.0036502438597381115, + 0.00593924755230546, + 0.003082871902734041, + 0.003527111141011119, + 0.00387102784588933, + 0.004673256538808346, + 0.00395327340811491, + 0.0035793487913906574, + 0.005440470762550831, + 0.003666170872747898, + 0.006483759731054306, + 0.0038521415553987026, + 0.043903499841690063, + 0.04218553006649017, + 0.035103876143693924, + 0.0328245647251606, + 0.07541732490062714, + 0.04665452614426613, + 0.05204426124691963, + 0.028437398374080658, + 0.007735808379948139, + 0.013970760628581047, + 0.012959438376128674, + 0.012130465358495712, + 0.016326947137713432, + 0.012846611440181732, + 0.010792369954288006, + 0.008514739573001862, + 0.01899871788918972, + 0.017277179285883904, + 0.01062035746872425, + 0.016805734485387802, + 0.02760530821979046, + 0.014890233054757118, + 0.05714163929224014, + 0.016201471909880638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.008635440841317177 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902975916862488 + ], + "opposite_logprob_tokens": [ + -2.4640297889709473 + ], + "logprob_reward": -0.08902975916862488, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902975916862488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17116272449493408, + 0.056895386427640915, + 0.04393154755234718, + 0.2851410210132599, + 0.0810910314321518, + 0.02350625768303871, + 0.031998664140701294, + 0.18589995801448822, + 0.058916568756103516, + 0.061456818133592606 + ], + "credit_map": [ + 0.018108047544956207, + 0.031857602298259735, + 0.020198484882712364, + 0.018142035230994225, + 0.020603466778993607, + 0.019019046798348427, + 0.024814195930957794, + 0.018419846892356873, + 0.006940807681530714, + 0.01267823576927185, + 0.00530566181987524, + 0.006302163004875183, + 0.007059745490550995, + 0.005085374228656292, + 0.007776767946779728, + 0.005746633280068636, + 0.005135640501976013, + 0.005729991011321545, + 0.004810386803001165, + 0.0043794624507427216, + 0.007418548688292503, + 0.004554924555122852, + 0.007602144498378038, + 0.004300450440496206, + 0.036041080951690674, + 0.029511859640479088, + 0.02327720820903778, + 0.02000199817121029, + 0.04490068554878235, + 0.03033130243420601, + 0.044574085623025894, + 0.056502822786569595, + 0.007812509313225746, + 0.011686996556818485, + 0.008763488382101059, + 0.01076328195631504, + 0.010786496102809906, + 0.00982065312564373, + 0.01177168171852827, + 0.00968592707067728, + 0.0024305928964167833, + 0.0033025590237230062, + 0.0034611253067851067, + 0.0021656365133821964, + 0.0027437633834779263, + 0.003722814377397299, + 0.002976472256705165, + 0.0027032927609980106, + 0.002716479357331991, + 0.0033229775726795197, + 0.0037719181273132563, + 0.0029794685542583466, + 0.005096526350826025, + 0.0036215726286172867, + 0.007114490494132042, + 0.0033752312883734703, + 0.02343970723450184, + 0.022989707067608833, + 0.01783689297735691, + 0.01450476422905922, + 0.0388496108353138, + 0.025780389085412025, + 0.027615200728178024, + 0.0148836774751544, + 0.00645810691639781, + 0.00797053799033165, + 0.00622155237942934, + 0.005430752877146006, + 0.01050766371190548, + 0.005793655291199684, + 0.01144739706069231, + 0.00508690532296896, + 0.004366835579276085, + 0.00591627974063158, + 0.006258342415094376, + 0.00929710641503334, + 0.008162400685250759, + 0.004696421790868044, + 0.01611584611237049, + 0.006643583998084068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.08902975916862488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007616763934493065 + ], + "opposite_logprob_tokens": [ + -4.882616996765137 + ], + "logprob_reward": -0.007616763934493065, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007616763934493065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12221153825521469, + 0.029003756120800972, + 0.033291399478912354, + 0.1286490112543106, + 0.038918137550354004, + 0.023463500663638115, + 0.028575748205184937, + 0.04473224654793739, + 0.06626054644584656, + 0.4848940670490265 + ], + "credit_map": [ + 0.015153784304857254, + 0.025660159066319466, + 0.018833359703421593, + 0.013394580222666264, + 0.014207135885953903, + 0.012439273297786713, + 0.01193952839821577, + 0.010583717375993729, + 0.0030258349142968655, + 0.003989820368587971, + 0.004341781139373779, + 0.0038836540188640356, + 0.0033824180718511343, + 0.0027150558307766914, + 0.0042045144364237785, + 0.0034606775734573603, + 0.004008416552096605, + 0.005343940574675798, + 0.0042059896513819695, + 0.0038748790975660086, + 0.0046151867136359215, + 0.003132822923362255, + 0.004600326996296644, + 0.0035098385997116566, + 0.015630614012479782, + 0.013468173332512379, + 0.011809774674475193, + 0.009183600544929504, + 0.020104577764868736, + 0.014332661405205727, + 0.019398434087634087, + 0.0247211791574955, + 0.0040813470259308815, + 0.005853327922523022, + 0.004676200915127993, + 0.0054688286036252975, + 0.00526601355522871, + 0.005045811180025339, + 0.004371823277324438, + 0.004154784604907036, + 0.002410504501312971, + 0.0027245194651186466, + 0.004561008419841528, + 0.0022517775651067495, + 0.0025227817241102457, + 0.002991199027746916, + 0.0029807095415890217, + 0.0030210018157958984, + 0.003495830809697509, + 0.0031445526983588934, + 0.004758183378726244, + 0.002801260445266962, + 0.004511286970227957, + 0.0031501867342740297, + 0.003982230555266142, + 0.002732217777520418, + 0.0037602728698402643, + 0.00485239876434207, + 0.005588590633124113, + 0.004124806262552738, + 0.007996706292033195, + 0.004590360447764397, + 0.009285937994718552, + 0.0045331730507314205, + 0.005899821873754263, + 0.00873715989291668, + 0.013942694291472435, + 0.008151458576321602, + 0.008143415674567223, + 0.00596588384360075, + 0.008818247355520725, + 0.006601864006370306, + 0.05134175345301628, + 0.03042326308786869, + 0.039062581956386566, + 0.05116870626807213, + 0.0396386943757534, + 0.024843845516443253, + 0.18298867344856262, + 0.06542655825614929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.007616763934493065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, sideways tracking shot captures the fluidity of Dunhuang murals, depicting an ancient carpet depicting mythical beasts, steadily and calmly navigating the air. Made of thick, handwoven wool, the carpet's aged texture is evident in its slightly frayed edges and crimson tassels that flutter backward in the wind. The carpet's primary colors are aged ochre and lapis lazuli blue, meticulously woven with bold yet fluid calligraphic lines to create an intricate pattern: a graceful, ornately winged nine-colored deer at its center, surrounded by ethereal flying fairies and stylized auspicious clouds, all shimmering with the characteristic matte texture of mineral pigments. The carpet, rather than moving in a rigid, lifelike motion, rises and falls steadily through the air, as if sailing on invisible waves. Below the camera, a vast expanse of golden desert stretches to the horizon. The setting sun paints the rolling dunes a warm, flowing sea of gold, casting long, soft blue-purple shadows on the leeward sides. The sky, far from a realistic depiction, is a splendid azure, punctuated by elongated, brushstroke-like orange-red clouds. The entire scene is bathed in warm and soft light, imbued with a sacred, tranquil, and epic atmosphere, as if a millennia-old mural has been brought to life and is slowly unfolding before our eyes.", + "seed": 6352, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_4/index_2386_seed_6352_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026075528608635067, + "vlm_logprob": -0.026075528608635067, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.026075528608635067, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01756434701383114 + ], + "opposite_logprob_tokens": [ + -4.142564296722412 + ], + "logprob_reward": -0.01756434701383114, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01756434701383114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14687128365039825, + 0.08427932858467102, + 0.0455516092479229, + 0.07119113951921463, + 0.033403802663087845, + 0.0648348331451416, + 0.24961014091968536, + 0.0885777547955513, + 0.17946386337280273, + 0.03621618077158928 + ], + "credit_map": [ + 0.007699431385844946, + 0.015589268878102303, + 0.010338012129068375, + 0.03576347231864929, + 0.019353443756699562, + 0.007468543481081724, + 0.038821980357170105, + 0.011837140657007694, + 0.0058966223150491714, + 0.014971679076552391, + 0.007379601243883371, + 0.010724538937211037, + 0.018779953941702843, + 0.006713458336889744, + 0.01159715186804533, + 0.00821632333099842, + 0.004710719920694828, + 0.007306166924536228, + 0.005176493898034096, + 0.006524255499243736, + 0.0069211432710289955, + 0.004404779989272356, + 0.00548582011833787, + 0.005022228695452213, + 0.005855453200638294, + 0.009124943986535072, + 0.0066702766343951225, + 0.010338612832129002, + 0.013639689423143864, + 0.0069864788092672825, + 0.008413693867623806, + 0.01016199216246605, + 0.003587382612749934, + 0.0051196543499827385, + 0.0035190500784665346, + 0.004985443782061338, + 0.00444706529378891, + 0.003848775289952755, + 0.00409183744341135, + 0.003804598469287157, + 0.004822490271180868, + 0.009792270138859749, + 0.0064674243330955505, + 0.009609983302652836, + 0.012892499566078186, + 0.007063091732561588, + 0.007781697902828455, + 0.006405378691852093, + 0.02431885153055191, + 0.021757423877716064, + 0.052574273198843, + 0.018397118896245956, + 0.03310307487845421, + 0.018831787630915642, + 0.05694981664419174, + 0.023677824065089226, + 0.009517035447061062, + 0.01078908983618021, + 0.013915193267166615, + 0.007519797421991825, + 0.015862150117754936, + 0.008112265728414059, + 0.014294063672423363, + 0.008568169549107552, + 0.024400528520345688, + 0.027672240510582924, + 0.02385079674422741, + 0.018981395289301872, + 0.019146151840686798, + 0.020118290558457375, + 0.017665231600403786, + 0.02762923762202263, + 0.004194350913167, + 0.00363555527292192, + 0.0066757225431501865, + 0.004169019404798746, + 0.0037738648243248463, + 0.0046866461634635925, + 0.005166049115359783, + 0.003914977423846722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.01756434701383114 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979971095919609 + ], + "opposite_logprob_tokens": [ + -3.529799699783325 + ], + "logprob_reward": -0.02979971095919609, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979971095919609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32262179255485535, + 0.04051435738801956, + 0.04268963634967804, + 0.03963175788521767, + 0.07961400598287582, + 0.040980104357004166, + 0.12024292349815369, + 0.11507277935743332, + 0.1734795868396759, + 0.025153102353215218 + ], + "credit_map": [ + 0.012628953903913498, + 0.018284285441040993, + 0.008845623582601547, + 0.1192268654704094, + 0.026973523199558258, + 0.0072920783422887325, + 0.11404028534889221, + 0.015330162830650806, + 0.003749886993318796, + 0.004843776114284992, + 0.0036869198083877563, + 0.00746740261092782, + 0.006043459288775921, + 0.0042391009628772736, + 0.006104662083089352, + 0.0043791490606963634, + 0.004619965795427561, + 0.004989341367036104, + 0.005418296437710524, + 0.006939091719686985, + 0.005095569416880608, + 0.00345348147675395, + 0.005255522206425667, + 0.006918369792401791, + 0.003380479756742716, + 0.006706784479320049, + 0.004001609515398741, + 0.00598130701109767, + 0.005033926106989384, + 0.0036701729986816645, + 0.005821545142680407, + 0.005035933572798967, + 0.014128686860203743, + 0.00883368868380785, + 0.009336673654615879, + 0.011498925276100636, + 0.009425296448171139, + 0.00854405015707016, + 0.007602254394441843, + 0.010244432836771011, + 0.004938072059303522, + 0.005763286259025335, + 0.004782562144100666, + 0.006424868479371071, + 0.006400418467819691, + 0.004205893725156784, + 0.004102229606360197, + 0.00436277175322175, + 0.015392723493278027, + 0.012511932291090488, + 0.01581168919801712, + 0.012365403585135937, + 0.01707448624074459, + 0.011503000743687153, + 0.01886608451604843, + 0.016717607155442238, + 0.013616351410746574, + 0.009948892518877983, + 0.018888575956225395, + 0.011406874284148216, + 0.014618748798966408, + 0.011588558554649353, + 0.016557572409510612, + 0.018447205424308777, + 0.024320734664797783, + 0.02629840560257435, + 0.022778719663619995, + 0.019025487825274467, + 0.018371975049376488, + 0.020045220851898193, + 0.017480943351984024, + 0.02515811286866665, + 0.0024095484986901283, + 0.002817995147779584, + 0.006022014655172825, + 0.001975125167518854, + 0.002362352330237627, + 0.0017158703412860632, + 0.003931218758225441, + 0.00391897838562727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.02979971095919609 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01106978952884674 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.01106978952884674, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01106978952884674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48015663027763367, + 0.0343889482319355, + 0.049931012094020844, + 0.07079073041677475, + 0.038297392427921295, + 0.02297915704548359, + 0.1310809999704361, + 0.08864964544773102, + 0.06193740665912628, + 0.02178814634680748 + ], + "credit_map": [ + 0.018375728279352188, + 0.026804380118846893, + 0.010243958793580532, + 0.1782827526330948, + 0.04687332361936569, + 0.010856046341359615, + 0.16701379418373108, + 0.021706635132431984, + 0.0029571843333542347, + 0.004937731195241213, + 0.0037681004032492638, + 0.004405288025736809, + 0.007283382583409548, + 0.0034530104603618383, + 0.003902267199009657, + 0.0036819817032665014, + 0.0036026835441589355, + 0.004900120198726654, + 0.004624628461897373, + 0.006257330533117056, + 0.01017004530876875, + 0.005057613365352154, + 0.008862403221428394, + 0.0064561814069747925, + 0.004726009909063578, + 0.009867751970887184, + 0.00510005047544837, + 0.01215016283094883, + 0.014545508660376072, + 0.00920072291046381, + 0.007572397124022245, + 0.007628118619322777, + 0.00512317568063736, + 0.004285736940801144, + 0.004395484458655119, + 0.006351102609187365, + 0.005151715595275164, + 0.004379525315016508, + 0.00458715483546257, + 0.0040234909392893314, + 0.0017082587582990527, + 0.003479906590655446, + 0.002222905633971095, + 0.0030581201426684856, + 0.0049698795191943645, + 0.00259197223931551, + 0.0029676377307623625, + 0.00198047561571002, + 0.009695432148873806, + 0.008847245946526527, + 0.03389901667833328, + 0.00848920363932848, + 0.016033027321100235, + 0.00749233178794384, + 0.03605702146887779, + 0.010567707009613514, + 0.011275341734290123, + 0.010571470484137535, + 0.009988763369619846, + 0.009158732369542122, + 0.016390608623623848, + 0.00963648222386837, + 0.012560341507196426, + 0.00906788744032383, + 0.008739149197936058, + 0.008197586052119732, + 0.00924624502658844, + 0.005861890967935324, + 0.00714620528742671, + 0.006305700168013573, + 0.006086324341595173, + 0.01035429909825325, + 0.0025537481997162104, + 0.0028584818355739117, + 0.0028121487703174353, + 0.0030476434621959925, + 0.0034332505892962217, + 0.0028356427792459726, + 0.002079922007396817, + 0.0021673073060810566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.01106978952884674 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000958919525146 + ], + "opposite_logprob_tokens": [ + -2.695009708404541 + ], + "logprob_reward": -0.07000958919525146, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000958919525146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2252153754234314, + 0.026152921840548515, + 0.023480337113142014, + 0.04761406406760216, + 0.040949705988168716, + 0.06148126348853111, + 0.19668954610824585, + 0.11916327476501465, + 0.19750992953777313, + 0.061743609607219696 + ], + "credit_map": [ + 0.009875395335257053, + 0.014351481571793556, + 0.007382252253592014, + 0.07844778150320053, + 0.020348254591226578, + 0.006425593048334122, + 0.076289601624012, + 0.012095044367015362, + 0.0026029101572930813, + 0.004828605800867081, + 0.0030357297509908676, + 0.0031302431598305702, + 0.004934058524668217, + 0.002116319490596652, + 0.00286850961856544, + 0.0026365441735833883, + 0.00236556981690228, + 0.003769013099372387, + 0.0022195696365088224, + 0.0028423762414604425, + 0.003949019126594067, + 0.002974887378513813, + 0.0025465860962867737, + 0.0028133157175034285, + 0.004906519316136837, + 0.005643055308610201, + 0.0054891170002520084, + 0.00629690894857049, + 0.0064749885350465775, + 0.003519283840432763, + 0.008102328516542912, + 0.007181860506534576, + 0.005119919776916504, + 0.004614950157701969, + 0.00488482927903533, + 0.006686613894999027, + 0.0053139906376600266, + 0.004733211360871792, + 0.004409214947372675, + 0.005186975933611393, + 0.005222782492637634, + 0.009355146437883377, + 0.0056659323163330555, + 0.010386068373918533, + 0.011771226301789284, + 0.007403707131743431, + 0.006116526201367378, + 0.005559868179261684, + 0.023638682439923286, + 0.018705245107412338, + 0.030885936692357063, + 0.018859459087252617, + 0.025328677147626877, + 0.019050350412726402, + 0.03627150133252144, + 0.023949701339006424, + 0.017011918127536774, + 0.012934574857354164, + 0.01509054098278284, + 0.012797803618013859, + 0.018926415592432022, + 0.012900355271995068, + 0.016452165320515633, + 0.013049487955868244, + 0.02896072342991829, + 0.029450029134750366, + 0.02530130371451378, + 0.020741697400808334, + 0.02130485512316227, + 0.022820191457867622, + 0.01988215558230877, + 0.029048975557088852, + 0.005967501550912857, + 0.0072633083909749985, + 0.014094796031713486, + 0.006293071433901787, + 0.005500514525920153, + 0.004669541493058205, + 0.008332212455570698, + 0.00962266605347395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.07000958919525146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a thick, handwoven wool carpet featuring mythical beasts and calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the carpet display aged ochre and lapis lazuli blue colors with crimson tassels fluttering backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central pattern a graceful, ornately winged nine-colored deer surrounded by flying fairies and stylized clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the carpet rise and fall steadily through the air with a matte texture of mineral pigments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast golden desert with a splendid azure sky containing brushstroke-like orange-red clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019342063460499048 + ], + "opposite_logprob_tokens": [ + -6.251934051513672 + ], + "logprob_reward": -0.0019342063460499048, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019342063460499048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14484941959381104, + 0.07571381330490112, + 0.12162560224533081, + 0.08688154071569443, + 0.08195934444665909, + 0.07101902365684509, + 0.13172514736652374, + 0.08629433065652847, + 0.15476329624652863, + 0.045168470591306686 + ], + "credit_map": [ + 0.013370221480727196, + 0.011294994503259659, + 0.012114848010241985, + 0.02417132444679737, + 0.019695378839969635, + 0.011577690951526165, + 0.027192573994398117, + 0.02543237805366516, + 0.008006113581359386, + 0.00937803927809, + 0.010538644157350063, + 0.008795182220637798, + 0.008966735564172268, + 0.007735066581517458, + 0.011704779230058193, + 0.010589259676635265, + 0.008795717731118202, + 0.01344822533428669, + 0.014867755584418774, + 0.012512028217315674, + 0.013297281228005886, + 0.01134649757295847, + 0.018240654841065407, + 0.029117446392774582, + 0.011078760027885437, + 0.008936326950788498, + 0.009619815275073051, + 0.011974653229117393, + 0.008690559305250645, + 0.009659898467361927, + 0.012721709907054901, + 0.014199819415807724, + 0.01327356044203043, + 0.008428688161075115, + 0.010566575452685356, + 0.01306649949401617, + 0.008368102833628654, + 0.008465240709483624, + 0.00833766721189022, + 0.011453017592430115, + 0.010096762329339981, + 0.007781289983540773, + 0.010612405836582184, + 0.008686400018632412, + 0.010890733450651169, + 0.005302848760038614, + 0.009085359051823616, + 0.008563226088881493, + 0.014303866773843765, + 0.010340030305087566, + 0.028669219464063644, + 0.01122734323143959, + 0.012606571428477764, + 0.010361394844949245, + 0.030109357088804245, + 0.0141073577105999, + 0.011650609783828259, + 0.00826768483966589, + 0.0141718415543437, + 0.010505309328436852, + 0.012071680277585983, + 0.00870584324002266, + 0.01175359170883894, + 0.009167775511741638, + 0.02205466665327549, + 0.02173677459359169, + 0.01792006380856037, + 0.016835711896419525, + 0.018108269199728966, + 0.01723339408636093, + 0.01627933979034424, + 0.024595079943537712, + 0.007411046419292688, + 0.004813785664737225, + 0.005043398588895798, + 0.004656339529901743, + 0.005080106668174267, + 0.005487400572746992, + 0.004706397652626038, + 0.007969996891915798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0019342063460499048 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_9/index_tensor([5499, 5499])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_9/index_tensor([5499, 5499])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6ee83f94e54cafa824b34503e04f08b72145d4e0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_9/index_tensor([5499, 5499])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a fixed, high-angle camera, the first rays of golden morning sunlight stream down from a low angle, illuminating the vast expanse of rice paddies. Tender green rice seedlings stand neatly in the flooded fields, the mirror-like surface of the water reflecting the warm orange-red sky. A thin layer of water vapor slowly rises from between the roots of the rice seedlings and the surface of the water, gathering into a wispy, translucent white mist that gently hovers above the paddy fields. As sunlight penetrates this mist, the light becomes soft and layered, giving the mist a subtle golden glow. Meanwhile, dewdrops condensed on the tips of the rice seedlings' emerald leaves reflect the radiance like diamonds. The entire scene exudes a profound tranquility and poetry, its primary color palette a vibrant green and warm gold creating a peaceful atmosphere imbued with vitality and hope.", + "seed": 23567, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_9/index_5499_seed_23567_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12710029927548022, + "vlm_logprob": -0.12710029927548022, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.12710029927548022, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05603129044175148 + ], + "opposite_logprob_tokens": [ + -3.3060312271118164 + ], + "logprob_reward": -0.05603129044175148, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05603129044175148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22007954120635986, + 0.13606151938438416, + 0.04975039139389992, + 0.09007143974304199, + 0.040586356073617935, + 0.031692102551460266, + 0.03903019428253174, + 0.04912414029240608, + 0.01986396685242653, + 0.3237404227256775 + ], + "credit_map": [ + 0.024986259639263153, + 0.016974179074168205, + 0.03912680968642235, + 0.03497376665472984, + 0.018857687711715698, + 0.015177274122834206, + 0.027401061728596687, + 0.042582474648952484, + 0.007722146809101105, + 0.008832667022943497, + 0.025410637259483337, + 0.04088791459798813, + 0.004803157877177, + 0.006856982596218586, + 0.015843700617551804, + 0.02570430375635624, + 0.00280676013790071, + 0.0037452278193086386, + 0.010495457798242569, + 0.00834782887250185, + 0.004722043871879578, + 0.00480098370462656, + 0.00832761824131012, + 0.006504470016807318, + 0.004519660025835037, + 0.019448136910796165, + 0.009822694584727287, + 0.009441875852644444, + 0.02296130545437336, + 0.008494002744555473, + 0.006807446014136076, + 0.008576314896345139, + 0.002410815330222249, + 0.007737616542726755, + 0.011464100331068039, + 0.0049896989949047565, + 0.003905615769326687, + 0.0036014411598443985, + 0.0034062338527292013, + 0.0030708317644894123, + 0.002261271234601736, + 0.0031968154944479465, + 0.007795440964400768, + 0.004044596571475267, + 0.0025637790095061064, + 0.0037872097454965115, + 0.0031975978054106236, + 0.004845390096306801, + 0.0027469731867313385, + 0.0033535833936184645, + 0.005563467741012573, + 0.007842891849577427, + 0.0024771043099462986, + 0.0037076466251164675, + 0.006878140848129988, + 0.006460385397076607, + 0.003650917671620846, + 0.003647803794592619, + 0.007933330722153187, + 0.007854197174310684, + 0.0034377656411379576, + 0.005012753419578075, + 0.00922437570989132, + 0.00836299080401659, + 0.0009614126174710691, + 0.0007854111026972532, + 0.0015747540164738894, + 0.0021399958059191704, + 0.0007956358604133129, + 0.0009777792729437351, + 0.006706323940306902, + 0.005922653246670961, + 0.0052832248620688915, + 0.002772811334580183, + 0.01430737879127264, + 0.02099010720849037, + 0.00656484579667449, + 0.006667522247880697, + 0.14350448548793793, + 0.12365001440048218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.05603129044175148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867476224899292 + ], + "opposite_logprob_tokens": [ + -3.0486748218536377 + ], + "logprob_reward": -0.04867476224899292, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867476224899292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10701730102300644, + 0.15494254231452942, + 0.10234593600034714, + 0.11570094525814056, + 0.03715560957789421, + 0.04389086365699768, + 0.09459319710731506, + 0.09174522757530212, + 0.0179910808801651, + 0.23461730778217316 + ], + "credit_map": [ + 0.009762527421116829, + 0.009862793609499931, + 0.01772853545844555, + 0.021735528483986855, + 0.00766874710097909, + 0.009176453575491905, + 0.011815800331532955, + 0.019266916438937187, + 0.009571458213031292, + 0.011531699448823929, + 0.029203996062278748, + 0.023704426363110542, + 0.008759697899222374, + 0.013449549674987793, + 0.025443747639656067, + 0.033277980983257294, + 0.006590250879526138, + 0.0074943737126886845, + 0.013352329842746258, + 0.028573542833328247, + 0.00754676666110754, + 0.007817215286195278, + 0.019328707829117775, + 0.011642747558653355, + 0.007297043222934008, + 0.012746802531182766, + 0.0206734798848629, + 0.021336393430829048, + 0.008567312732338905, + 0.015597362071275711, + 0.016938332468271255, + 0.012544223107397556, + 0.002219164976850152, + 0.007046561222523451, + 0.01012580655515194, + 0.004566397052258253, + 0.002711889101192355, + 0.002720683114603162, + 0.00488365488126874, + 0.002881451975554228, + 0.0027886831667274237, + 0.0023092220071703196, + 0.013795135542750359, + 0.006387366447597742, + 0.0026074545457959175, + 0.0032367899548262358, + 0.0049936664290726185, + 0.007772548124194145, + 0.006810890510678291, + 0.006413455121219158, + 0.014014683663845062, + 0.022317636758089066, + 0.005585446022450924, + 0.007961533963680267, + 0.01689952239394188, + 0.014590022154152393, + 0.007892214693129063, + 0.005216098390519619, + 0.013791938312351704, + 0.021472692489624023, + 0.005037378519773483, + 0.004883573390543461, + 0.020048581063747406, + 0.013402756303548813, + 0.0013306906912475824, + 0.0014561318093910813, + 0.0026106887962669134, + 0.005334502086043358, + 0.001099014887586236, + 0.0014464107807725668, + 0.002525900723412633, + 0.0021877402905374765, + 0.003954985179007053, + 0.002571167191490531, + 0.010280825197696686, + 0.01570763811469078, + 0.005045904312282801, + 0.004894273355603218, + 0.10357001423835754, + 0.08859249949455261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.04867476224899292 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054957568645477295 + ], + "opposite_logprob_tokens": [ + -2.929957628250122 + ], + "logprob_reward": -0.054957568645477295, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054957568645477295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11524941027164459, + 0.04763258248567581, + 0.08027073740959167, + 0.07068319618701935, + 0.02578268200159073, + 0.033054448664188385, + 0.09380614757537842, + 0.04523579776287079, + 0.023374779149889946, + 0.46491026878356934 + ], + "credit_map": [ + 0.01114612352102995, + 0.01185250747948885, + 0.01704578287899494, + 0.02811235748231411, + 0.008821087889373302, + 0.008268053643405437, + 0.013222410343587399, + 0.016781093552708626, + 0.002718965522944927, + 0.0034486991353332996, + 0.009353317320346832, + 0.012547687627375126, + 0.0025040218606591225, + 0.003672015154734254, + 0.007869463413953781, + 0.005518412217497826, + 0.0047535886988043785, + 0.005340311676263809, + 0.012356825172901154, + 0.02015179954469204, + 0.009890965186059475, + 0.007770181633532047, + 0.012116906233131886, + 0.00789016205817461, + 0.0033243559300899506, + 0.0070743318647146225, + 0.015907088294625282, + 0.012578255496919155, + 0.010352648794651031, + 0.0069512189365923405, + 0.00870251189917326, + 0.005792786832898855, + 0.0014617235865443945, + 0.004600845742970705, + 0.007778432220220566, + 0.004238881636410952, + 0.001639324938878417, + 0.001470120158046484, + 0.0022728191688656807, + 0.002320533152669668, + 0.0019859883468598127, + 0.0025350982323288918, + 0.009823666885495186, + 0.00595613569021225, + 0.0020565190352499485, + 0.002671279711648822, + 0.0038892081938683987, + 0.0041365548968315125, + 0.008741425350308418, + 0.007813907228410244, + 0.018932120874524117, + 0.017304612323641777, + 0.007085597608238459, + 0.0069779809564352036, + 0.012167098000645638, + 0.014783401973545551, + 0.0035406474489718676, + 0.003390308003872633, + 0.005696533247828484, + 0.007343886885792017, + 0.0037294134963303804, + 0.004649918060749769, + 0.009515260346233845, + 0.0073698293417692184, + 0.0011662663891911507, + 0.0007745492039248347, + 0.0014897385844960809, + 0.0018536570714786649, + 0.0009723996627144516, + 0.0012051919475197792, + 0.008612369187176228, + 0.007300608325749636, + 0.0075696236453950405, + 0.003806806169450283, + 0.020444391295313835, + 0.030366331338882446, + 0.009603474289178848, + 0.009179594926536083, + 0.20649747550487518, + 0.1774425357580185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.054957568645477295 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743257761001587 + ], + "opposite_logprob_tokens": [ + -0.9743257761001587 + ], + "logprob_reward": -0.4743257761001587, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743257761001587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0452859029173851, + 0.05849205702543259, + 0.05474160611629486, + 0.03491342067718506, + 0.023271847516298294, + 0.02846776694059372, + 0.08695059269666672, + 0.06138212978839874, + 0.017700906842947006, + 0.5887936949729919 + ], + "credit_map": [ + 0.005092390812933445, + 0.0056724767200648785, + 0.007796016987413168, + 0.007013299502432346, + 0.005058562383055687, + 0.0032850769348442554, + 0.0054178726859390736, + 0.005950210615992546, + 0.0037050710525363684, + 0.0036507239565253258, + 0.012110360898077488, + 0.008608921431005001, + 0.0029491393361240625, + 0.0046030376106500626, + 0.009401469491422176, + 0.013463336043059826, + 0.0031861073803156614, + 0.003932344727218151, + 0.010466582141816616, + 0.012336916290223598, + 0.007026142906397581, + 0.0061019170098006725, + 0.006601649336516857, + 0.005089954473078251, + 0.002618849743157625, + 0.00535526592284441, + 0.005018388852477074, + 0.004506224766373634, + 0.004651319235563278, + 0.005225008353590965, + 0.00399146368727088, + 0.0035469024442136288, + 0.001298605464398861, + 0.00207985145971179, + 0.0038196600507944822, + 0.0037576963659375906, + 0.0025780913420021534, + 0.0020103410352021456, + 0.004727131221443415, + 0.003000472206622362, + 0.0018483385210856795, + 0.0019609949085861444, + 0.007248878479003906, + 0.004162509460002184, + 0.001958634704351425, + 0.002238950692117214, + 0.0034945467486977577, + 0.005554917734116316, + 0.007137187756597996, + 0.009608728811144829, + 0.015648042783141136, + 0.016649354249238968, + 0.006355112884193659, + 0.007650251500308514, + 0.012805111706256866, + 0.011096806265413761, + 0.004826561082154512, + 0.004649100825190544, + 0.009713796898722649, + 0.013609718531370163, + 0.004233182407915592, + 0.003617299720644951, + 0.012324643321335316, + 0.00840783677995205, + 0.0013289459748193622, + 0.0015260978834703565, + 0.001787885557860136, + 0.0021426575258374214, + 0.0012595057487487793, + 0.0014892134349793196, + 0.004770106170326471, + 0.0033964957110583782, + 0.009261341765522957, + 0.005085617303848267, + 0.025478670373558998, + 0.03722202032804489, + 0.012144537642598152, + 0.011957934126257896, + 0.26097261905670166, + 0.22667096555233002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.4743257761001587 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015120989410206676 + ], + "opposite_logprob_tokens": [ + -6.501512050628662 + ], + "logprob_reward": -0.0015120989410206676, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015120989410206676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06446453183889389, + 0.04451632872223854, + 0.055399090051651, + 0.04319725185632706, + 0.03151778131723404, + 0.012809650972485542, + 0.025067990645766258, + 0.05593619868159294, + 0.008237853646278381, + 0.6588533520698547 + ], + "credit_map": [ + 0.007178529165685177, + 0.00571080669760704, + 0.011184818111360073, + 0.010851171799004078, + 0.00515861576423049, + 0.003578955540433526, + 0.009527647867798805, + 0.011273985728621483, + 0.0029487244319170713, + 0.0031339058186858892, + 0.010080774314701557, + 0.007498015649616718, + 0.0028551339637488127, + 0.0031103254295885563, + 0.005996656604111195, + 0.008892793208360672, + 0.0029225845355540514, + 0.0037824460305273533, + 0.011809444054961205, + 0.013247465714812279, + 0.005973778665065765, + 0.002888027112931013, + 0.010005900636315346, + 0.004769442602992058, + 0.0023026319686323404, + 0.003493078052997589, + 0.010462024249136448, + 0.009430593810975552, + 0.003992362879216671, + 0.0028045405633747578, + 0.006093935575336218, + 0.004618083126842976, + 0.0017647772328928113, + 0.004353135824203491, + 0.009133673273026943, + 0.0051974402740597725, + 0.0020074034109711647, + 0.0015852717915549874, + 0.004258002154529095, + 0.0032180773559957743, + 0.000955660012550652, + 0.0014789962442591786, + 0.0012780155520886183, + 0.0020568338222801685, + 0.001120006083510816, + 0.0014715961879119277, + 0.002283665584400296, + 0.002164877485483885, + 0.002010749187320471, + 0.0021800969261676073, + 0.004795367363840342, + 0.0046730428002774715, + 0.0016534131718799472, + 0.0022021308541297913, + 0.0032413541339337826, + 0.0043118358589708805, + 0.003654101165011525, + 0.0037346123717725277, + 0.00996658205986023, + 0.013349151238799095, + 0.003670445177704096, + 0.003798775840550661, + 0.010646611452102661, + 0.007115921005606651, + 0.0006442858721129596, + 0.0005794446915388107, + 0.000930076465010643, + 0.0019213524647057056, + 0.0006297245272435248, + 0.0006818489055149257, + 0.0015887415502220392, + 0.0012623787624761462, + 0.010757836513221264, + 0.005405504256486893, + 0.028799068182706833, + 0.042619600892066956, + 0.013474637642502785, + 0.012980020605027676, + 0.2925012409687042, + 0.2523154318332672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.0015120989410206676 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a fixed, high-angle camera, the first rays of golden morning sunlight stream down from a low angle, illuminating the vast expanse of rice paddies. Tender green rice seedlings stand neatly in the flooded fields, the mirror-like surface of the water reflecting the warm orange-red sky. A thin layer of water vapor slowly rises from between the roots of the rice seedlings and the surface of the water, gathering into a wispy, translucent white mist that gently hovers above the paddy fields. As sunlight penetrates this mist, the light becomes soft and layered, giving the mist a subtle golden glow. Meanwhile, dewdrops condensed on the tips of the rice seedlings' emerald leaves reflect the radiance like diamonds. The entire scene exudes a profound tranquility and poetry, its primary color palette a vibrant green and warm gold creating a peaceful atmosphere imbued with vitality and hope.", + "seed": 23568, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_9/index_5499_seed_23568_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06761889366898685, + "vlm_logprob": -0.06761889366898685, + "vlm_yes_no_margin": 4.124999952316284, + "vlm_token_credit": -0.06761889366898685, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0516965277493 + ], + "opposite_logprob_tokens": [ + -4.051696300506592 + ], + "logprob_reward": -0.0516965277493, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0516965277493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06154714152216911, + 0.07588104158639908, + 0.04252871125936508, + 0.08227607607841492, + 0.12467872351408005, + 0.16479451954364777, + 0.19868126511573792, + 0.06952159106731415, + 0.03927714750170708, + 0.14081384241580963 + ], + "credit_map": [ + 0.006781308446079493, + 0.013450429774820805, + 0.009025247767567635, + 0.008229164406657219, + 0.005539204925298691, + 0.009953679516911507, + 0.003039704170078039, + 0.005528395064175129, + 0.013935773633420467, + 0.009727145545184612, + 0.007790963631123304, + 0.004874011501669884, + 0.014945019036531448, + 0.01339710969477892, + 0.005215287674218416, + 0.005995721556246281, + 0.008475501090288162, + 0.006277807522565126, + 0.004631277173757553, + 0.006229731719940901, + 0.007411233149468899, + 0.004491555038839579, + 0.0022550032008439302, + 0.0027565984055399895, + 0.013576318509876728, + 0.020457876846194267, + 0.010993757285177708, + 0.011000615544617176, + 0.006968376692384481, + 0.0074639227241277695, + 0.004864344373345375, + 0.006950852461159229, + 0.02089817449450493, + 0.009903145022690296, + 0.008438710123300552, + 0.006574507802724838, + 0.021925916895270348, + 0.04727931693196297, + 0.00348023883998394, + 0.006178708281368017, + 0.03048205003142357, + 0.06808691471815109, + 0.006544433068484068, + 0.009640893898904324, + 0.014301596209406853, + 0.02577994205057621, + 0.004032030701637268, + 0.0059266467578709126, + 0.013691681437194347, + 0.0580684058368206, + 0.015048407949507236, + 0.029696354642510414, + 0.04624259099364281, + 0.016949402168393135, + 0.009172547608613968, + 0.009811854921281338, + 0.008673966862261295, + 0.0061184074729681015, + 0.00477487500756979, + 0.005563139915466309, + 0.022117627784609795, + 0.015799086540937424, + 0.0032273209653794765, + 0.0032471551094204187, + 0.003974127117544413, + 0.0043084947392344475, + 0.0073111774399876595, + 0.005386514589190483, + 0.008697913959622383, + 0.004969999194145203, + 0.0021655678283423185, + 0.002463350770995021, + 0.022076688706874847, + 0.012831936590373516, + 0.010825141333043575, + 0.009388744831085205, + 0.04765382409095764, + 0.02486257255077362, + 0.005864811595529318, + 0.007310114800930023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.0516965277493 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006748029962182045 + ], + "opposite_logprob_tokens": [ + -5.006748199462891 + ], + "logprob_reward": -0.006748029962182045, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006748029962182045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12353452295064926, + 0.11949808150529861, + 0.09119822829961777, + 0.1318928450345993, + 0.06538812071084976, + 0.07328605651855469, + 0.16509240865707397, + 0.11728498339653015, + 0.034051503986120224, + 0.07877331972122192 + ], + "credit_map": [ + 0.026375416666269302, + 0.015950439497828484, + 0.008542787283658981, + 0.008988196030259132, + 0.014118046499788761, + 0.03733070194721222, + 0.004366008564829826, + 0.00786292739212513, + 0.022444773465394974, + 0.012910068966448307, + 0.009577088057994843, + 0.010158427059650421, + 0.01448223926126957, + 0.03600084409117699, + 0.00554052134975791, + 0.00838410947471857, + 0.011454492807388306, + 0.017439857125282288, + 0.004573829937726259, + 0.006604821886867285, + 0.025105375796556473, + 0.020223719999194145, + 0.00288171274587512, + 0.002914419397711754, + 0.021686749532818794, + 0.03254137560725212, + 0.014376121573150158, + 0.0173067394644022, + 0.010488664731383324, + 0.016605103388428688, + 0.007306471932679415, + 0.011581621132791042, + 0.010745328851044178, + 0.009353073313832283, + 0.00783186499029398, + 0.007279696874320507, + 0.006645468529313803, + 0.01562855765223503, + 0.003141614841297269, + 0.004762518219649792, + 0.013292056508362293, + 0.028683556243777275, + 0.0036160629242658615, + 0.0035561337135732174, + 0.006986579392105341, + 0.010730608366429806, + 0.0027244808152318, + 0.0036965743638575077, + 0.014142455533146858, + 0.05405649170279503, + 0.008568267337977886, + 0.012195938266813755, + 0.053540900349617004, + 0.013785232789814472, + 0.004354560747742653, + 0.004448567051440477, + 0.007137592416256666, + 0.008495745249092579, + 0.005996945314109325, + 0.0057075973600149155, + 0.038299258798360825, + 0.041677143424749374, + 0.004408662673085928, + 0.005562029778957367, + 0.003259677905589342, + 0.0049109854735434055, + 0.0041648889891803265, + 0.00480703404173255, + 0.006592517253011465, + 0.004573632497340441, + 0.002685283310711384, + 0.0030574831180274487, + 0.008591586723923683, + 0.006619926542043686, + 0.00677812285721302, + 0.008214354515075684, + 0.026264213025569916, + 0.013788360171020031, + 0.004521156661212444, + 0.003995595965534449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.006748029962182045 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347704231739044 + ], + "opposite_logprob_tokens": [ + -3.6513476371765137 + ], + "logprob_reward": -0.026347704231739044, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347704231739044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09680457413196564, + 0.09312482178211212, + 0.09708506613969803, + 0.14025045931339264, + 0.10495565831661224, + 0.0710824504494667, + 0.16720834374427795, + 0.18762414157390594, + 0.015129108913242817, + 0.026735428720712662 + ], + "credit_map": [ + 0.011514324694871902, + 0.015578478574752808, + 0.017422717064619064, + 0.0183574091643095, + 0.007231318391859531, + 0.013928291387856007, + 0.005230873357504606, + 0.007541165221482515, + 0.019516268745064735, + 0.008429376408457756, + 0.0063476720824837685, + 0.005727472715079784, + 0.012279580347239971, + 0.0292658731341362, + 0.004492643289268017, + 0.007065923884510994, + 0.006272478029131889, + 0.018538113683462143, + 0.005042273085564375, + 0.007589663378894329, + 0.03294072300195694, + 0.020036613568663597, + 0.0029053459875285625, + 0.003759847255423665, + 0.02375871315598488, + 0.030338222160935402, + 0.018304916098713875, + 0.01825539767742157, + 0.011407075449824333, + 0.014633248560130596, + 0.009195915423333645, + 0.014356961473822594, + 0.01577301137149334, + 0.009745306335389614, + 0.008118728175759315, + 0.006604220252484083, + 0.0184672512114048, + 0.03737447038292885, + 0.0032054660841822624, + 0.005667190533131361, + 0.014988408423960209, + 0.027799353003501892, + 0.0041951509192585945, + 0.0035574701614677906, + 0.007198965176939964, + 0.008216865360736847, + 0.001925421180203557, + 0.0032008232083171606, + 0.023853015154600143, + 0.06731940060853958, + 0.010219196788966656, + 0.009095381014049053, + 0.035123102366924286, + 0.011809748597443104, + 0.004037376027554274, + 0.0057511115446686745, + 0.01914397068321705, + 0.02524503879249096, + 0.013228461146354675, + 0.01758299581706524, + 0.04179980233311653, + 0.04869972541928291, + 0.007799671962857246, + 0.014124464243650436, + 0.001962450332939625, + 0.002729802392423153, + 0.002086223801597953, + 0.0024380965624004602, + 0.0013642879202961922, + 0.0016866362420842052, + 0.0012456184485927224, + 0.0016159925144165754, + 0.0025924970395863056, + 0.002305190311744809, + 0.002221399452537298, + 0.00330062722787261, + 0.009052720852196217, + 0.004602960776537657, + 0.0014003330143168569, + 0.0012596985325217247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.026347704231739044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521229684352875 + ], + "opposite_logprob_tokens": [ + -1.5021229982376099 + ], + "logprob_reward": -0.2521229684352875, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521229684352875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1297047883272171, + 0.07099219411611557, + 0.0547541081905365, + 0.0566459484398365, + 0.1228112056851387, + 0.14719533920288086, + 0.0507943369448185, + 0.16825291514396667, + 0.053141284734010696, + 0.1457078903913498 + ], + "credit_map": [ + 0.019770868122577667, + 0.029391910880804062, + 0.01491000410169363, + 0.02100258506834507, + 0.010068496689200401, + 0.021640976890921593, + 0.006205922923982143, + 0.006714021787047386, + 0.011397618800401688, + 0.010854305699467659, + 0.00606129365041852, + 0.00491580693051219, + 0.014146692119538784, + 0.014006051234900951, + 0.004759267903864384, + 0.004851158242672682, + 0.006772586610168219, + 0.007398521061986685, + 0.0037716557271778584, + 0.0032762615010142326, + 0.019040733575820923, + 0.0092200031504035, + 0.0021799933165311813, + 0.0030943574383854866, + 0.009972833096981049, + 0.010563653893768787, + 0.006931936834007502, + 0.006414882373064756, + 0.005762169603258371, + 0.005877962335944176, + 0.004376912489533424, + 0.006745596881955862, + 0.022312108427286148, + 0.012258386239409447, + 0.009519034065306187, + 0.007491725962609053, + 0.019325535744428635, + 0.04076352342963219, + 0.004394058603793383, + 0.006746836472302675, + 0.03087795339524746, + 0.06118644401431084, + 0.006537932436913252, + 0.0074137425981462, + 0.015707403421401978, + 0.015600564889609814, + 0.003964414354413748, + 0.005906874313950539, + 0.0063631185330450535, + 0.006091372109949589, + 0.005312210880219936, + 0.006444845348596573, + 0.012688558548688889, + 0.007943786680698395, + 0.002656326163560152, + 0.0032941203098744154, + 0.01480169128626585, + 0.016294853761792183, + 0.020234785974025726, + 0.017412874847650528, + 0.04218030348420143, + 0.042068466544151306, + 0.00577678345143795, + 0.009483149275183678, + 0.006571418605744839, + 0.007995354011654854, + 0.009173023514449596, + 0.006570346187800169, + 0.006709669250994921, + 0.006234188564121723, + 0.0045105330646038055, + 0.005376750137656927, + 0.016017021611332893, + 0.010584883391857147, + 0.012859695591032505, + 0.01316134911030531, + 0.045559465885162354, + 0.02474684827029705, + 0.011523817665874958, + 0.011254812590777874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.2521229684352875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011792379664257169 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.0011792379664257169, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011792379664257169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13521595299243927, + 0.16155864298343658, + 0.03969211131334305, + 0.16715848445892334, + 0.0497538186609745, + 0.05851992219686508, + 0.11042463779449463, + 0.20488665997982025, + 0.031323596835136414, + 0.041466135531663895 + ], + "credit_map": [ + 0.026563677936792374, + 0.018953688442707062, + 0.011685558594763279, + 0.009650909341871738, + 0.015229973010718822, + 0.03752722963690758, + 0.006914730649441481, + 0.00869019515812397, + 0.021279089152812958, + 0.02481890842318535, + 0.015826933085918427, + 0.01728946901857853, + 0.024955635890364647, + 0.03311771899461746, + 0.009622927755117416, + 0.014647958800196648, + 0.004565357230603695, + 0.006673682481050491, + 0.005708142649382353, + 0.007404245436191559, + 0.004544905852526426, + 0.005286084488034248, + 0.0022057932801544666, + 0.0033039015252143145, + 0.03180798888206482, + 0.040193647146224976, + 0.018620172515511513, + 0.020520493388175964, + 0.01354659628123045, + 0.015801701694726944, + 0.010720143094658852, + 0.01594776101410389, + 0.007353490684181452, + 0.00584263214841485, + 0.004616098012775183, + 0.004368734546005726, + 0.007786280009895563, + 0.014088168740272522, + 0.0022280453704297543, + 0.003470372874289751, + 0.010232066735625267, + 0.007095920387655497, + 0.005661942530423403, + 0.005975221283733845, + 0.007406218443065882, + 0.012858822010457516, + 0.0037621064111590385, + 0.00552763044834137, + 0.013394325971603394, + 0.04421943053603172, + 0.0050734058022499084, + 0.005982703994959593, + 0.025883670896291733, + 0.008692906238138676, + 0.002761244308203459, + 0.004416957031935453, + 0.026800893247127533, + 0.015705810859799385, + 0.020710838958621025, + 0.010890381410717964, + 0.06284134835004807, + 0.05120560899376869, + 0.007527448236942291, + 0.009204353205859661, + 0.003911616280674934, + 0.003838134929537773, + 0.005168817937374115, + 0.003150970907881856, + 0.004463414195924997, + 0.0035780034959316254, + 0.0029064249247312546, + 0.004306215792894363, + 0.005935965571552515, + 0.004798990674316883, + 0.004321932327002287, + 0.004066284280270338, + 0.011019277386367321, + 0.006080801598727703, + 0.002170129679143429, + 0.0030727561097592115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.0011792379664257169 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a fixed, high-angle camera, the first rays of golden morning sunlight stream down from a low angle, illuminating the vast expanse of rice paddies. Tender green rice seedlings stand neatly in the flooded fields, the mirror-like surface of the water reflecting the warm orange-red sky. A thin layer of water vapor slowly rises from between the roots of the rice seedlings and the surface of the water, gathering into a wispy, translucent white mist that gently hovers above the paddy fields. As sunlight penetrates this mist, the light becomes soft and layered, giving the mist a subtle golden glow. Meanwhile, dewdrops condensed on the tips of the rice seedlings' emerald leaves reflect the radiance like diamonds. The entire scene exudes a profound tranquility and poetry, its primary color palette a vibrant green and warm gold creating a peaceful atmosphere imbued with vitality and hope.", + "seed": 23569, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_9/index_5499_seed_23569_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07292781821452081, + "vlm_logprob": -0.07292781821452081, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.07292781821452081, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1283154934644699 + ], + "opposite_logprob_tokens": [ + -2.2533154487609863 + ], + "logprob_reward": -0.1283154934644699, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1283154934644699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20825952291488647, + 0.16765373945236206, + 0.108220174908638, + 0.05672749504446983, + 0.059101857244968414, + 0.02656249888241291, + 0.2583371102809906, + 0.028793253004550934, + 0.028073281049728394, + 0.05827109515666962 + ], + "credit_map": [ + 0.01209789328277111, + 0.013858086429536343, + 0.029318489134311676, + 0.025981120765209198, + 0.024267781525850296, + 0.02115560509264469, + 0.04444042965769768, + 0.037140123546123505, + 0.01664111204445362, + 0.013109537772834301, + 0.02347651869058609, + 0.03296079486608505, + 0.013167058117687702, + 0.01761389523744583, + 0.036739282310009, + 0.013945541344583035, + 0.009944912046194077, + 0.015205569565296173, + 0.010693000629544258, + 0.010416891425848007, + 0.008667084388434887, + 0.010310791432857513, + 0.02553677000105381, + 0.01744515635073185, + 0.0039986069314181805, + 0.003198081161826849, + 0.00777076743543148, + 0.013757212087512016, + 0.0039098006673157215, + 0.011825443245470524, + 0.007930980063974857, + 0.004336602985858917, + 0.003249074798077345, + 0.0035513867624104023, + 0.007966334000229836, + 0.013261662796139717, + 0.004114645998924971, + 0.005188894923776388, + 0.016507310792803764, + 0.005262549966573715, + 0.0026515736244618893, + 0.003244393737986684, + 0.0031573947053402662, + 0.003054664935916662, + 0.003837698372080922, + 0.004092167131602764, + 0.004129758104681969, + 0.0023948487360030413, + 0.025675030425190926, + 0.02526845782995224, + 0.030354708433151245, + 0.033473387360572815, + 0.04508308693766594, + 0.03381320834159851, + 0.037229180335998535, + 0.027440037578344345, + 0.002485024742782116, + 0.0037461163010448217, + 0.0026654747780412436, + 0.0047689517959952354, + 0.0034873494878411293, + 0.003595213871449232, + 0.004839455708861351, + 0.003205663990229368, + 0.003105764277279377, + 0.003935745917260647, + 0.0035630601923912764, + 0.003160337219014764, + 0.003129330463707447, + 0.002772475825622678, + 0.004527973476797342, + 0.003878593910485506, + 0.005122530739754438, + 0.010703321546316147, + 0.003945564851164818, + 0.002654516836628318, + 0.006005451548844576, + 0.019550424069166183, + 0.0067415619269013405, + 0.003547726199030876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.1283154934644699 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650686591863632 + ], + "opposite_logprob_tokens": [ + -4.758650779724121 + ], + "logprob_reward": -0.008650686591863632, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650686591863632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2308618426322937, + 0.05509626120328903, + 0.10874246805906296, + 0.11541017889976501, + 0.17357204854488373, + 0.03455498069524765, + 0.045030467212200165, + 0.061683788895606995, + 0.043391160666942596, + 0.13165682554244995 + ], + "credit_map": [ + 0.018491175025701523, + 0.02349518984556198, + 0.023919686675071716, + 0.025895502418279648, + 0.0475832037627697, + 0.03256409615278244, + 0.02676747366786003, + 0.03214551508426666, + 0.0054978360421955585, + 0.003730092430487275, + 0.00908277090638876, + 0.008005852811038494, + 0.0067115286365151405, + 0.007620207034051418, + 0.009384719654917717, + 0.005063255317509174, + 0.01818183995783329, + 0.03273523226380348, + 0.006241003517061472, + 0.006742782425135374, + 0.008351203985512257, + 0.022699721157550812, + 0.006813636049628258, + 0.006977052893489599, + 0.005191202741116285, + 0.005248187109827995, + 0.013616562820971012, + 0.03472287952899933, + 0.00692353630438447, + 0.029163947328925133, + 0.012417756952345371, + 0.008126107975840569, + 0.01340864971280098, + 0.01696745678782463, + 0.026676712557673454, + 0.02365083619952202, + 0.017904574051499367, + 0.021103933453559875, + 0.03439083322882652, + 0.019469065591692924, + 0.0034184372052550316, + 0.004180505871772766, + 0.004066525958478451, + 0.004345052409917116, + 0.004392239265143871, + 0.005065044853836298, + 0.005251341033726931, + 0.0038358354941010475, + 0.004416222684085369, + 0.005580639000982046, + 0.004890646785497665, + 0.008055935613811016, + 0.00704077398404479, + 0.0058466470800340176, + 0.004777975380420685, + 0.004421625752002001, + 0.005361577030271292, + 0.006363024935126305, + 0.005974242463707924, + 0.006234051659703255, + 0.006830075290054083, + 0.006708182860165834, + 0.01440347544848919, + 0.00980916153639555, + 0.004614746198058128, + 0.005709792021661997, + 0.00515652634203434, + 0.004834786523133516, + 0.00567822577431798, + 0.005541608203202486, + 0.00583826145157218, + 0.006017211824655533, + 0.006061910185962915, + 0.012753186747431755, + 0.009718267247080803, + 0.00714095588773489, + 0.013740232214331627, + 0.04502188041806221, + 0.012027031742036343, + 0.025193361565470695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.008650686591863632 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328656055033207 + ], + "opposite_logprob_tokens": [ + -3.7732865810394287 + ], + "logprob_reward": -0.02328656055033207, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328656055033207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1330062448978424, + 0.2740788161754608, + 0.12491648644208908, + 0.0908811166882515, + 0.18022668361663818, + 0.03710894286632538, + 0.05022018775343895, + 0.034233253449201584, + 0.014136207289993763, + 0.061191968619823456 + ], + "credit_map": [ + 0.017670268192887306, + 0.01576915755867958, + 0.01576194167137146, + 0.014246618375182152, + 0.020609285682439804, + 0.015732698142528534, + 0.009046966210007668, + 0.02416931837797165, + 0.02174256183207035, + 0.017717331647872925, + 0.01480718981474638, + 0.07832537591457367, + 0.016041718423366547, + 0.022107956930994987, + 0.09022343903779984, + 0.013113269582390785, + 0.020688258111476898, + 0.03610727936029434, + 0.008716121315956116, + 0.008283406496047974, + 0.009336229413747787, + 0.02159152738749981, + 0.01116958912461996, + 0.009024086408317089, + 0.007547412067651749, + 0.008141000755131245, + 0.010386958718299866, + 0.020628321915864944, + 0.0078069791197776794, + 0.018742933869361877, + 0.011521861888468266, + 0.006105647422373295, + 0.010973596014082432, + 0.012687301263213158, + 0.014603884890675545, + 0.03943784907460213, + 0.01603558287024498, + 0.016507957130670547, + 0.05345393717288971, + 0.016526592895388603, + 0.003693803446367383, + 0.005169947165995836, + 0.003693132195621729, + 0.004595744889229536, + 0.005814176052808762, + 0.00609894935041666, + 0.004374003503471613, + 0.003669189987704158, + 0.004546695854514837, + 0.004919400904327631, + 0.0071677458472549915, + 0.005690861493349075, + 0.009554925374686718, + 0.0072160521522164345, + 0.00583276990801096, + 0.005291744135320187, + 0.003177113365381956, + 0.00510507682338357, + 0.0034381337463855743, + 0.0034562579821795225, + 0.005423952825367451, + 0.005327905062586069, + 0.0041725547052919865, + 0.004132261499762535, + 0.001114754006266594, + 0.0015688104322180152, + 0.0015696530463173985, + 0.0015999490860849619, + 0.0021341671235859394, + 0.0022097029723227024, + 0.002386614680290222, + 0.0015525565249845386, + 0.00513077899813652, + 0.0067825764417648315, + 0.006851745769381523, + 0.004706351552158594, + 0.008111675269901752, + 0.01760507933795452, + 0.00573755195364356, + 0.006266218610107899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.02328656055033207 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015710175037384 + ], + "opposite_logprob_tokens": [ + -1.701570987701416 + ], + "logprob_reward": -0.2015710175037384, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015710175037384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11050980538129807, + 0.12097819894552231, + 0.12447154521942139, + 0.02417253516614437, + 0.2496880739927292, + 0.022096429020166397, + 0.035015884786844254, + 0.047826942056417465, + 0.022146733477711678, + 0.243093803524971 + ], + "credit_map": [ + 0.008909142576158047, + 0.008766832761466503, + 0.01417494472116232, + 0.011778884567320347, + 0.01950501836836338, + 0.01069041807204485, + 0.021000349894165993, + 0.01568421721458435, + 0.012821538373827934, + 0.011139442212879658, + 0.010290277190506458, + 0.027410846203565598, + 0.00932404026389122, + 0.011357207782566547, + 0.03200441226363182, + 0.006630437448620796, + 0.019835129380226135, + 0.03444883972406387, + 0.008033456280827522, + 0.008710537105798721, + 0.011141078546643257, + 0.022295402362942696, + 0.011125156655907631, + 0.0088819470256567, + 0.00203151348978281, + 0.0019487052923068404, + 0.005738541949540377, + 0.003215585835278034, + 0.0014910866739228368, + 0.002585815265774727, + 0.003956781234592199, + 0.003204507287591696, + 0.016044098883867264, + 0.017130171880126, + 0.024967694655060768, + 0.05124538391828537, + 0.02067721262574196, + 0.023461347445845604, + 0.07378587871789932, + 0.02237629145383835, + 0.0020037400536239147, + 0.00205871369689703, + 0.0032601708080619574, + 0.0028052297420799732, + 0.0028094628360122442, + 0.003166461829096079, + 0.0037796932738274336, + 0.0022129579447209835, + 0.0028264184948056936, + 0.0031415619887411594, + 0.005033621098846197, + 0.004828476347029209, + 0.006060321815311909, + 0.004469063598662615, + 0.004025177098810673, + 0.004631245508790016, + 0.006094188429415226, + 0.006442829500883818, + 0.0040966845117509365, + 0.004706012550741434, + 0.007149528246372938, + 0.007896866649389267, + 0.006020530126988888, + 0.0054203057661652565, + 0.0026451696176081896, + 0.003409815486520529, + 0.002584389178082347, + 0.0026103011332452297, + 0.0030767854768782854, + 0.002990735461935401, + 0.002462164731696248, + 0.0023673721589148045, + 0.016243254765868187, + 0.0362444669008255, + 0.013344175182282925, + 0.010000545531511307, + 0.02629878558218479, + 0.09717098623514175, + 0.009347617626190186, + 0.03444398194551468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.2015710175037384 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028153329622000456 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.0028153329622000456, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028153329622000456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18435201048851013, + 0.15739686787128448, + 0.04117700085043907, + 0.07661663740873337, + 0.08700736612081528, + 0.058325186371803284, + 0.02415998838841915, + 0.049377720803022385, + 0.06092037260532379, + 0.2606668174266815 + ], + "credit_map": [ + 0.014590618200600147, + 0.012733165174722672, + 0.018548274412751198, + 0.02625848911702633, + 0.02666998654603958, + 0.024545960128307343, + 0.032422348856925964, + 0.028583182021975517, + 0.011813443154096603, + 0.010241108015179634, + 0.01193692721426487, + 0.042983319610357285, + 0.009279673919081688, + 0.014113927260041237, + 0.048811174929142, + 0.00821730401366949, + 0.0039852699264883995, + 0.007781603839248419, + 0.003246084088459611, + 0.0035765802022069693, + 0.0025782755110412836, + 0.009771712124347687, + 0.004644798114895821, + 0.0055926828645169735, + 0.005700340494513512, + 0.005570206791162491, + 0.010928276926279068, + 0.016088491305708885, + 0.00555587001144886, + 0.013543608598411083, + 0.010304379276931286, + 0.008925480768084526, + 0.0079092001542449, + 0.009078853763639927, + 0.012944486923515797, + 0.012333718128502369, + 0.008131122216582298, + 0.008868744596838951, + 0.016947416588664055, + 0.010793836787343025, + 0.00637459009885788, + 0.011017399840056896, + 0.0043266937136650085, + 0.005417007487267256, + 0.005529625806957483, + 0.01519799791276455, + 0.006173398345708847, + 0.004288480617105961, + 0.0024561716709285975, + 0.0020784588996320963, + 0.002370076486840844, + 0.0028064954094588757, + 0.004479572176933289, + 0.003147748066112399, + 0.003922736272215843, + 0.002898732665926218, + 0.005409139208495617, + 0.006160259246826172, + 0.004217206966131926, + 0.004914331715553999, + 0.009334948845207691, + 0.006794331129640341, + 0.007414359599351883, + 0.00513314688578248, + 0.0052068596705794334, + 0.009396243840456009, + 0.004998691380023956, + 0.004880597814917564, + 0.007771956268697977, + 0.01418263465166092, + 0.006693641189485788, + 0.007789749652147293, + 0.017923502251505852, + 0.040287308394908905, + 0.014236514456570148, + 0.009876479394733906, + 0.028096698224544525, + 0.10694413632154465, + 0.01157032698392868, + 0.03173189237713814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0028153329622000456 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a fixed, high-angle camera, the first rays of golden morning sunlight stream down from a low angle, illuminating the vast expanse of rice paddies. Tender green rice seedlings stand neatly in the flooded fields, the mirror-like surface of the water reflecting the warm orange-red sky. A thin layer of water vapor slowly rises from between the roots of the rice seedlings and the surface of the water, gathering into a wispy, translucent white mist that gently hovers above the paddy fields. As sunlight penetrates this mist, the light becomes soft and layered, giving the mist a subtle golden glow. Meanwhile, dewdrops condensed on the tips of the rice seedlings' emerald leaves reflect the radiance like diamonds. The entire scene exudes a profound tranquility and poetry, its primary color palette a vibrant green and warm gold creating a peaceful atmosphere imbued with vitality and hope.", + "seed": 23570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000093/group_9/index_5499_seed_23570_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05494442497147247, + "vlm_logprob": -0.05494442497147247, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.05494442497147247, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03741290792822838 + ], + "opposite_logprob_tokens": [ + -3.7874128818511963 + ], + "logprob_reward": -0.03741290792822838, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03741290792822838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3274466395378113, + 0.08479533344507217, + 0.026862436905503273, + 0.06104300171136856, + 0.040958572179079056, + 0.059880368411540985, + 0.10410477221012115, + 0.1319170445203781, + 0.07045487314462662, + 0.09253696352243423 + ], + "credit_map": [ + 0.03635237365961075, + 0.029643794521689415, + 0.04587601125240326, + 0.02787967026233673, + 0.054015401750802994, + 0.05327335000038147, + 0.030574575066566467, + 0.04983147233724594, + 0.010484986938536167, + 0.013011149130761623, + 0.0102187879383564, + 0.007513702381402254, + 0.014198065735399723, + 0.012415062636137009, + 0.0063966140151023865, + 0.010556971654295921, + 0.0043487101793289185, + 0.00424287561327219, + 0.0025755506940186024, + 0.002960250712931156, + 0.004618627950549126, + 0.002634006552398205, + 0.0029422345105558634, + 0.0025401818566024303, + 0.008073177188634872, + 0.009235942736268044, + 0.008197392337024212, + 0.006489929743111134, + 0.0069343033246695995, + 0.00644462276250124, + 0.008852973580360413, + 0.006814669352024794, + 0.004572506062686443, + 0.005234938580542803, + 0.0038724751211702824, + 0.0033320379443466663, + 0.006424095947295427, + 0.007500376086682081, + 0.0031811348162591457, + 0.006841013673692942, + 0.010086174122989178, + 0.008444733917713165, + 0.006232393439859152, + 0.004988366737961769, + 0.006512236315757036, + 0.007306178566068411, + 0.006644585635513067, + 0.00966570433229208, + 0.021774282678961754, + 0.015070914290845394, + 0.006744314916431904, + 0.005585148464888334, + 0.011529560200870037, + 0.014455736614763737, + 0.009756526909768581, + 0.01918829418718815, + 0.03024296648800373, + 0.02762410417199135, + 0.01280965842306614, + 0.010209351778030396, + 0.010790204629302025, + 0.01409460511058569, + 0.010755658149719238, + 0.015390504151582718, + 0.01034204475581646, + 0.009890754707157612, + 0.005159309599548578, + 0.0033631811384111643, + 0.013329553417861462, + 0.01679201051592827, + 0.004267718642950058, + 0.007310306653380394, + 0.016178833320736885, + 0.019964534789323807, + 0.0076401084661483765, + 0.009717429988086224, + 0.009448508732020855, + 0.010144316591322422, + 0.007163375150412321, + 0.012279867194592953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.03741290792822838 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649504743516445 + ], + "opposite_logprob_tokens": [ + -4.7586493492126465 + ], + "logprob_reward": -0.008649504743516445, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649504743516445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22648993134498596, + 0.09145449101924896, + 0.09224987030029297, + 0.07448156923055649, + 0.02488691173493862, + 0.062468256801366806, + 0.07013693451881409, + 0.08692552149295807, + 0.04734259843826294, + 0.22356389462947845 + ], + "credit_map": [ + 0.024539628997445107, + 0.021459857001900673, + 0.026086410507559776, + 0.029183046892285347, + 0.034810200333595276, + 0.03087446466088295, + 0.027168579399585724, + 0.03236774355173111, + 0.010662872344255447, + 0.01031795609742403, + 0.011180015280842781, + 0.008884633891284466, + 0.010757986456155777, + 0.010622851550579071, + 0.013244434259831905, + 0.01578373648226261, + 0.010792052373290062, + 0.00979545246809721, + 0.009753476828336716, + 0.00774043332785368, + 0.013643060810863972, + 0.015879886224865913, + 0.010037488304078579, + 0.01460802648216486, + 0.008294242434203625, + 0.007189061492681503, + 0.007600210607051849, + 0.00627236207947135, + 0.012026431038975716, + 0.013651910237967968, + 0.008201752789318562, + 0.011245588771998882, + 0.0034299222752451897, + 0.003207031637430191, + 0.004217114765197039, + 0.002755198162049055, + 0.002819635206833482, + 0.001925142016261816, + 0.0036817293148487806, + 0.002851137425750494, + 0.008402064442634583, + 0.008779327385127544, + 0.006988024804741144, + 0.00492407102137804, + 0.00640625786036253, + 0.009050092659890652, + 0.00876200757920742, + 0.009156406857073307, + 0.009951459243893623, + 0.009918414987623692, + 0.007582534104585648, + 0.006355075165629387, + 0.006322778761386871, + 0.011636598967015743, + 0.006116870325058699, + 0.012253198772668839, + 0.014931336045265198, + 0.012974926270544529, + 0.00765024870634079, + 0.007695455104112625, + 0.0074918936006724834, + 0.013746880926191807, + 0.008420943282544613, + 0.014013825915753841, + 0.007115510758012533, + 0.008312826044857502, + 0.005904561839997768, + 0.004787530284374952, + 0.004882759414613247, + 0.005328109487891197, + 0.0047516124323010445, + 0.006259686313569546, + 0.04017019271850586, + 0.04343203082680702, + 0.01785864681005478, + 0.02217784710228443, + 0.019956035539507866, + 0.02383466437458992, + 0.02132759988307953, + 0.03480685502290726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.008649504743516445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343639940023422 + ], + "opposite_logprob_tokens": [ + -3.651343584060669 + ], + "logprob_reward": -0.026343639940023422, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343639940023422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21331635117530823, + 0.0644158124923706, + 0.06091216951608658, + 0.10089755058288574, + 0.05655800551176071, + 0.022349463775753975, + 0.07628101110458374, + 0.19356241822242737, + 0.03143993020057678, + 0.1802673190832138 + ], + "credit_map": [ + 0.02393418177962303, + 0.03310367837548256, + 0.01898060366511345, + 0.025694388896226883, + 0.03937409818172455, + 0.023853890597820282, + 0.03204941749572754, + 0.016326086595654488, + 0.007882950827479362, + 0.008668129332363605, + 0.00812776479870081, + 0.007968740537762642, + 0.008526651188731194, + 0.00816727802157402, + 0.0076739895157516, + 0.00740031199529767, + 0.006975922267884016, + 0.007953789085149765, + 0.0058426037430763245, + 0.0056402236223220825, + 0.010511724278330803, + 0.010828102938830853, + 0.005988217890262604, + 0.007171586621552706, + 0.010247002355754375, + 0.012175030075013638, + 0.008280221372842789, + 0.009188602678477764, + 0.0182000994682312, + 0.01961386576294899, + 0.010249260812997818, + 0.012943463400006294, + 0.00839998759329319, + 0.009985508397221565, + 0.010628572665154934, + 0.004481474868953228, + 0.00699347909539938, + 0.0053369090892374516, + 0.003607840044423938, + 0.007124232593923807, + 0.0029150296468287706, + 0.0032439613714814186, + 0.0045133326202631, + 0.001942869508638978, + 0.002342468360438943, + 0.002240631030872464, + 0.002985923318192363, + 0.0021652481518685818, + 0.00951821357011795, + 0.014789322391152382, + 0.009227964095771313, + 0.006644978187978268, + 0.008270235732197762, + 0.012993155978620052, + 0.006904746405780315, + 0.007932392880320549, + 0.04199082404375076, + 0.05086715146899223, + 0.017473649233579636, + 0.01291628833860159, + 0.020661940798163414, + 0.023470908403396606, + 0.010796458460390568, + 0.015385184437036514, + 0.0038182977586984634, + 0.0041624754667282104, + 0.004269598051905632, + 0.004050320480018854, + 0.0035486735869199038, + 0.004644506610929966, + 0.0037344698794186115, + 0.0032115885987877846, + 0.0313497819006443, + 0.03329030051827431, + 0.01621243916451931, + 0.017092594876885414, + 0.01680302806198597, + 0.021162506192922592, + 0.018433598801493645, + 0.025923067703843117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.026343639940023422 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159976184368134 + ], + "opposite_logprob_tokens": [ + -1.7015997171401978 + ], + "logprob_reward": -0.20159976184368134, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159976184368134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16613440215587616, + 0.06421129405498505, + 0.061900682747364044, + 0.036097366362810135, + 0.02604236640036106, + 0.03782571107149124, + 0.07756265252828598, + 0.22186575829982758, + 0.07337523996829987, + 0.23498453199863434 + ], + "credit_map": [ + 0.019881796091794968, + 0.01934654824435711, + 0.019365867599844933, + 0.019012201577425003, + 0.027175970375537872, + 0.02123899944126606, + 0.024579040706157684, + 0.015533984638750553, + 0.007314297370612621, + 0.006428562104701996, + 0.00757453590631485, + 0.007085764780640602, + 0.008959236554801464, + 0.007769804447889328, + 0.009144467301666737, + 0.009934622794389725, + 0.008066807873547077, + 0.008106732740998268, + 0.010555660352110863, + 0.006341049913316965, + 0.007895002141594887, + 0.008705971762537956, + 0.006753712426871061, + 0.0054757483303546906, + 0.0045941476710140705, + 0.0037345278542488813, + 0.004733354784548283, + 0.00411719735711813, + 0.004690921399742365, + 0.0039009866304695606, + 0.005376349203288555, + 0.004949879366904497, + 0.002610089024528861, + 0.0025165346451103687, + 0.0038020024076104164, + 0.0028716460801661015, + 0.0029784860089421272, + 0.00406718859449029, + 0.0025823484174907207, + 0.004614072386175394, + 0.006533890496939421, + 0.005109115969389677, + 0.0047020879574120045, + 0.003906483296304941, + 0.0047222631983459, + 0.0040617953054606915, + 0.0045242528431117535, + 0.004265822935849428, + 0.014192552305758, + 0.01049213856458664, + 0.008661474101245403, + 0.005868957843631506, + 0.006039455998688936, + 0.010338342748582363, + 0.006743737496435642, + 0.01522599533200264, + 0.046535417437553406, + 0.056763000786304474, + 0.020565617829561234, + 0.014773955568671227, + 0.023618340492248535, + 0.02727828174829483, + 0.012722193263471127, + 0.01960894465446472, + 0.009215131402015686, + 0.009568614885210991, + 0.0115578044205904, + 0.009931976906955242, + 0.008130841888487339, + 0.005276000592857599, + 0.008679871447384357, + 0.011014998890459538, + 0.04304696246981621, + 0.04416823014616966, + 0.021098364144563675, + 0.024239322170615196, + 0.022166907787322998, + 0.023279977962374687, + 0.023717639967799187, + 0.03326711058616638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.20159976184368134 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a fixed, high-angle viewpoint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tender green rice seedlings visible standing in flooded fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a thin layer of rising water vapor forming a wispy white mist above the fields?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dewdrops on the tips of the rice seedlings reflect light like diamonds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the primary color palette of the scene vibrant green and warm gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007163104019127786 + ], + "opposite_logprob_tokens": [ + -7.250716209411621 + ], + "logprob_reward": -0.0007163104019127786, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007163104019127786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10603035241365433, + 0.04616595059633255, + 0.05643199011683464, + 0.029663359746336937, + 0.01743302494287491, + 0.055611997842788696, + 0.15074720978736877, + 0.32322660088539124, + 0.026313139125704765, + 0.18837641179561615 + ], + "credit_map": [ + 0.013187803328037262, + 0.018322082236409187, + 0.0116896852850914, + 0.010442833416163921, + 0.013406000100076199, + 0.009931906126439571, + 0.015511753037571907, + 0.013538278639316559, + 0.006693299859762192, + 0.007195696234703064, + 0.004941535647958517, + 0.00497937249019742, + 0.00571393221616745, + 0.005426767747849226, + 0.004191545769572258, + 0.007023797836154699, + 0.0052660019136965275, + 0.005541871301829815, + 0.004908826667815447, + 0.0034937981981784105, + 0.01062101311981678, + 0.013260981999337673, + 0.004453903064131737, + 0.008885589428246021, + 0.003336684312671423, + 0.003937290515750647, + 0.0023252167738974094, + 0.0024459483101963997, + 0.004950350150465965, + 0.006203400902450085, + 0.0026041679084300995, + 0.0038602985441684723, + 0.002606490161269903, + 0.0034955909941345453, + 0.002640755381435156, + 0.0018790006870403886, + 0.00163166131824255, + 0.0017661378951743245, + 0.001914621563628316, + 0.001498767058365047, + 0.011642752215266228, + 0.009461150504648685, + 0.00455981120467186, + 0.0045203352347016335, + 0.0043805683963000774, + 0.005405811592936516, + 0.007499039638787508, + 0.00814252533018589, + 0.026149572804570198, + 0.029136264696717262, + 0.01320608425885439, + 0.01058227475732565, + 0.01641000062227249, + 0.0231451578438282, + 0.013611014932394028, + 0.018506821244955063, + 0.07138203084468842, + 0.08514945209026337, + 0.02920854277908802, + 0.023588264361023903, + 0.031012633815407753, + 0.035578370094299316, + 0.021427897736430168, + 0.0258793868124485, + 0.0037143928930163383, + 0.004396897740662098, + 0.0034483191557228565, + 0.0031911435071378946, + 0.0031752525828778744, + 0.0033466387540102005, + 0.002350773196667433, + 0.002689720829948783, + 0.03270146623253822, + 0.03663745895028114, + 0.017956390976905823, + 0.017685266211628914, + 0.01682707481086254, + 0.01901264861226082, + 0.017613990232348442, + 0.0299421064555645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.0007163104019127786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_1/index_tensor([9846, 9846])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_1/index_tensor([9846, 9846])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3b7f1e08bc1900e0cbd651cebd08e282c03775a5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_1/index_tensor([9846, 9846])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle close-up begins at the top of an inclined plane, as the camera begins a smooth, continuous downward tilt, carefully following the motion of a small toy car. The subject is a sleek, vibrant red plastic toy car, its simple molded body reflecting the ambient light, resting on an inclined unfinished pine plank, its grain clearly visible. The car is released from complete rest and begins its descent; its black plastic wheels begin to spin, accelerating rapidly and decisively across the wooden surface under the pull of gravity, quickly blurring into smooth motion. The camera maintains close tracking, capturing the seamless transition as the car rolls from the end of the ramp onto the polished light brown hardwood floor. Its momentum carries it forward across the flat surface, where it gradually slows, gliding silently for a few feet before coming to a gentle, complete stop. The entire scene is bathed in soft, warm interior light, creating subtle highlights on the car's plastic shell and the polished floor, establishing a clean, focused, and tranquil atmosphere that isolates the simple, elegant physics of motion.", + "seed": 5195, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_1/index_9846_seed_5195_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.006379666586872191, + "vlm_logprob": -0.006379666586872191, + "vlm_yes_no_margin": 5.55, + "vlm_token_credit": -0.006379666586872191, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005610315129160881 + ], + "opposite_logprob_tokens": [ + -5.505610466003418 + ], + "logprob_reward": -0.005610315129160881, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005610315129160881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09822693467140198, + 0.1227811649441719, + 0.1078619658946991, + 0.054499030113220215, + 0.10006076842546463, + 0.13830940425395966, + 0.2649860680103302, + 0.03493576496839523, + 0.03436503931879997, + 0.04397385194897652 + ], + "credit_map": [ + 0.013455113396048546, + 0.020071400329470634, + 0.014911112375557423, + 0.009269029833376408, + 0.009180684573948383, + 0.014367864467203617, + 0.006837361957877874, + 0.010134371928870678, + 0.010630120523273945, + 0.0162128284573555, + 0.01833507977426052, + 0.013821590691804886, + 0.017878366634249687, + 0.017701419070363045, + 0.016211071982979774, + 0.011990683153271675, + 0.010886725969612598, + 0.019749267026782036, + 0.021945323795080185, + 0.014300080016255379, + 0.012027191929519176, + 0.012303364463150501, + 0.008400564081966877, + 0.008249446749687195, + 0.004775980021804571, + 0.0066203163005411625, + 0.005766620393842459, + 0.010609986260533333, + 0.00711281830444932, + 0.008405309170484543, + 0.006938583683222532, + 0.004269415978342295, + 0.015214680694043636, + 0.013903839513659477, + 0.014390047639608383, + 0.011388488113880157, + 0.008908095769584179, + 0.009100658819079399, + 0.010458273813128471, + 0.016696682199835777, + 0.010714656673371792, + 0.014234675094485283, + 0.018297310918569565, + 0.015031290240585804, + 0.01924465037882328, + 0.027389144524931908, + 0.01286568958312273, + 0.02053198777139187, + 0.024774404242634773, + 0.021379953250288963, + 0.03390755504369736, + 0.02372918464243412, + 0.02732725255191326, + 0.06071101129055023, + 0.018849221989512444, + 0.0543074831366539, + 0.0030174818821251392, + 0.0028402998577803373, + 0.00588882016018033, + 0.0065683904103934765, + 0.003384773852303624, + 0.004969354718923569, + 0.004527112934738398, + 0.003739531384781003, + 0.002536315703764558, + 0.004630547948181629, + 0.004196752794086933, + 0.005023862235248089, + 0.004381891340017319, + 0.0035731482785195112, + 0.005525588057935238, + 0.004496933426707983, + 0.0066220141015946865, + 0.003406988922506571, + 0.0048772008158266544, + 0.006299822125583887, + 0.004229625221341848, + 0.004346434958279133, + 0.006874673534184694, + 0.007317089941352606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.005610315129160881 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648204617202282 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.008648204617202282, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648204617202282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.151504784822464, + 0.13765908777713776, + 0.08962714672088623, + 0.05230003222823143, + 0.040736258029937744, + 0.12519149482250214, + 0.1250152289867401, + 0.04938850179314613, + 0.1122441440820694, + 0.11633330583572388 + ], + "credit_map": [ + 0.019516393542289734, + 0.024061843752861023, + 0.024765679612755775, + 0.02133716270327568, + 0.012865145690739155, + 0.022427964955568314, + 0.013105551712214947, + 0.013425040058791637, + 0.02174173854291439, + 0.023848092183470726, + 0.0232431348413229, + 0.022512078285217285, + 0.01127105113118887, + 0.013362823985517025, + 0.0116715207695961, + 0.010008648037910461, + 0.011145154945552349, + 0.016285747289657593, + 0.013172900304198265, + 0.01208866573870182, + 0.011424052529036999, + 0.011361604556441307, + 0.008698972873389721, + 0.00545004615560174, + 0.006221469957381487, + 0.007342739496380091, + 0.005451150704175234, + 0.007980508729815483, + 0.0066710226237773895, + 0.0077136424370110035, + 0.006999034434556961, + 0.003920465242117643, + 0.0062000565230846405, + 0.00777240376919508, + 0.008015680126845837, + 0.005894842557609081, + 0.0036434917710721493, + 0.003028041450306773, + 0.003428966272622347, + 0.002752777189016342, + 0.01077075395733118, + 0.012474310584366322, + 0.006514632608741522, + 0.006142724305391312, + 0.006363104563206434, + 0.04397357255220413, + 0.004782716743648052, + 0.03416968137025833, + 0.012322062626481056, + 0.007600728422403336, + 0.014720752835273743, + 0.007427992299199104, + 0.007652638480067253, + 0.034611113369464874, + 0.008015884086489677, + 0.032664068043231964, + 0.005505031906068325, + 0.0037264577113091946, + 0.006660868879407644, + 0.005578020587563515, + 0.005244381260126829, + 0.010589545592665672, + 0.006119898520410061, + 0.005964295472949743, + 0.006718069780617952, + 0.012618274427950382, + 0.008709628134965897, + 0.0159140657633543, + 0.01340529229491949, + 0.011337507516145706, + 0.02148483134806156, + 0.02205648273229599, + 0.021813929080963135, + 0.009816732257604599, + 0.01190164778381586, + 0.015405213460326195, + 0.011739637702703476, + 0.010267537087202072, + 0.016260182484984398, + 0.01912842132151127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.008648204617202282 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189893379807472 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.014189893379807472, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189893379807472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0921909436583519, + 0.044516097754240036, + 0.08751107007265091, + 0.07815101742744446, + 0.07059313356876373, + 0.12591326236724854, + 0.3099059760570526, + 0.1479099690914154, + 0.019891340285539627, + 0.023417169228196144 + ], + "credit_map": [ + 0.012061363086104393, + 0.01696990802884102, + 0.017457090318202972, + 0.011829975992441177, + 0.007246042601764202, + 0.01091037131845951, + 0.007348638493567705, + 0.008367550559341908, + 0.006400248967111111, + 0.005532352719455957, + 0.006006764713674784, + 0.008256698027253151, + 0.005156456958502531, + 0.005687975324690342, + 0.004293335601687431, + 0.0031822659075260162, + 0.011666672304272652, + 0.017598848789930344, + 0.01472143828868866, + 0.011253587901592255, + 0.009273199364542961, + 0.007792473305016756, + 0.00871811993420124, + 0.0064867339096963406, + 0.008436985313892365, + 0.009029798209667206, + 0.008565058000385761, + 0.012254955247044563, + 0.01085394062101841, + 0.01253030076622963, + 0.009578127413988113, + 0.006901855580508709, + 0.01135284174233675, + 0.012582920491695404, + 0.014858279377222061, + 0.009474066086113453, + 0.005953563842922449, + 0.004666424822062254, + 0.005616183392703533, + 0.006088852882385254, + 0.013923660852015018, + 0.014410761184990406, + 0.008535433560609818, + 0.008855820633471012, + 0.008869118057191372, + 0.02913815528154373, + 0.007736489176750183, + 0.03444384038448334, + 0.03111136145889759, + 0.02591208927333355, + 0.03852362185716629, + 0.02351417765021324, + 0.029931137338280678, + 0.07460729777812958, + 0.022043289616703987, + 0.06426297128200531, + 0.011980886571109295, + 0.014112577773630619, + 0.031063104048371315, + 0.033061712980270386, + 0.014567370526492596, + 0.01563211902976036, + 0.01583056151866913, + 0.011661635711789131, + 0.00160651421174407, + 0.0022102564107626677, + 0.0031532063148915768, + 0.0036258779000490904, + 0.001817944459617138, + 0.0020685959607362747, + 0.0033722238149493933, + 0.0020367209799587727, + 0.003086310578510165, + 0.0035240829456597567, + 0.0029204373713582754, + 0.002653942909091711, + 0.0035085787530988455, + 0.002311836928129196, + 0.002684946171939373, + 0.0027270333375781775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.014189893379807472 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815927378833294 + ], + "opposite_logprob_tokens": [ + -5.8778157234191895 + ], + "logprob_reward": -0.002815927378833294, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815927378833294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11795191466808319, + 0.06824015080928802, + 0.03647870942950249, + 0.08753632009029388, + 0.07142075896263123, + 0.270633339881897, + 0.1718001663684845, + 0.06392544507980347, + 0.0760066956281662, + 0.036006540060043335 + ], + "credit_map": [ + 0.022890787571668625, + 0.023920586332678795, + 0.012116370722651482, + 0.014773058705031872, + 0.00719236908480525, + 0.010708052664995193, + 0.012008937075734138, + 0.014341752044856548, + 0.00808914378285408, + 0.014206038787961006, + 0.013214386999607086, + 0.008166216313838959, + 0.006099062506109476, + 0.007573799695819616, + 0.0061281598173081875, + 0.004763342905789614, + 0.005504301283508539, + 0.0062543838284909725, + 0.005064678844064474, + 0.004712976515293121, + 0.006062764208763838, + 0.003680030582472682, + 0.002699949312955141, + 0.0024996257852762938, + 0.008851404301822186, + 0.008063772693276405, + 0.011142878793179989, + 0.015122679062187672, + 0.012185214087367058, + 0.015568860806524754, + 0.01067810133099556, + 0.005923399701714516, + 0.015418070368468761, + 0.011755086481571198, + 0.016583293676376343, + 0.007684690877795219, + 0.004897537175565958, + 0.0037244237028062344, + 0.005108610261231661, + 0.006249052006751299, + 0.020461322739720345, + 0.020388759672641754, + 0.009597725234925747, + 0.010655246675014496, + 0.015729570761322975, + 0.09698645770549774, + 0.013277019374072552, + 0.08353723585605621, + 0.009655608795583248, + 0.009827598929405212, + 0.016153397038578987, + 0.012026979587972164, + 0.013197989203035831, + 0.051901303231716156, + 0.009921767748892307, + 0.04911553114652634, + 0.005578818265348673, + 0.005794881377369165, + 0.015528958290815353, + 0.011677873320877552, + 0.005138056818395853, + 0.005468990188091993, + 0.00843252893537283, + 0.006305340211838484, + 0.0073414575308561325, + 0.010253150947391987, + 0.013530747964978218, + 0.011653585359454155, + 0.0070964265614748, + 0.00883607566356659, + 0.009398692287504673, + 0.007896561175584793, + 0.005235603079199791, + 0.0038442646618932486, + 0.004838199354708195, + 0.004314952064305544, + 0.004244356416165829, + 0.004258262924849987, + 0.004482606425881386, + 0.004788298159837723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.002815927378833294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000633992429357022 + ], + "opposite_logprob_tokens": [ + -7.37563419342041 + ], + "logprob_reward": -0.000633992429357022, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.000633992429357022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17451688647270203, + 0.044515106827020645, + 0.09944820404052734, + 0.046798791736364365, + 0.05913536995649338, + 0.1082017794251442, + 0.17151238024234772, + 0.19097302854061127, + 0.06570051610469818, + 0.039197973906993866 + ], + "credit_map": [ + 0.027170686051249504, + 0.030453136190772057, + 0.03408072143793106, + 0.024366293102502823, + 0.011381878517568111, + 0.016841648146510124, + 0.017424114048480988, + 0.012798396870493889, + 0.005386725068092346, + 0.007950968109071255, + 0.010347526520490646, + 0.006075127515941858, + 0.0037275920622050762, + 0.004156933166086674, + 0.003380279988050461, + 0.003489956958219409, + 0.011544776149094105, + 0.018626682460308075, + 0.01769109256565571, + 0.013299490325152874, + 0.013844246044754982, + 0.010116157121956348, + 0.007944054901599884, + 0.006381699815392494, + 0.0052997637540102005, + 0.00467508053407073, + 0.004691667389124632, + 0.007601241581141949, + 0.006239050999283791, + 0.007910014130175114, + 0.005810656119138002, + 0.004571317695081234, + 0.008388983085751534, + 0.009697413071990013, + 0.012682177126407623, + 0.006686743814498186, + 0.004762199707329273, + 0.004148995969444513, + 0.004649011418223381, + 0.00811984483152628, + 0.009809977374970913, + 0.01103821862488985, + 0.007583468221127987, + 0.008682229556143284, + 0.009929873049259186, + 0.029678767547011375, + 0.007064642384648323, + 0.024414610117673874, + 0.03108372911810875, + 0.021192757412791252, + 0.028298623859882355, + 0.0174124576151371, + 0.022278297692537308, + 0.01678820513188839, + 0.018003633245825768, + 0.016454674303531647, + 0.012700887396931648, + 0.017098018899559975, + 0.0362793430685997, + 0.03742605447769165, + 0.019890744239091873, + 0.023718254640698433, + 0.02561100758612156, + 0.01824871450662613, + 0.005345017183572054, + 0.007122092414647341, + 0.012538952752947807, + 0.011900591664016247, + 0.006040917709469795, + 0.006341285537928343, + 0.010677307844161987, + 0.0057343486696481705, + 0.0040220539085567, + 0.00433049350976944, + 0.005025773774832487, + 0.006313137710094452, + 0.004098793491721153, + 0.003153157187625766, + 0.006394254043698311, + 0.005860313773155212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.000633992429357022 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle close-up begins at the top of an inclined plane, as the camera begins a smooth, continuous downward tilt, carefully following the motion of a small toy car. The subject is a sleek, vibrant red plastic toy car, its simple molded body reflecting the ambient light, resting on an inclined unfinished pine plank, its grain clearly visible. The car is released from complete rest and begins its descent; its black plastic wheels begin to spin, accelerating rapidly and decisively across the wooden surface under the pull of gravity, quickly blurring into smooth motion. The camera maintains close tracking, capturing the seamless transition as the car rolls from the end of the ramp onto the polished light brown hardwood floor. Its momentum carries it forward across the flat surface, where it gradually slows, gliding silently for a few feet before coming to a gentle, complete stop. The entire scene is bathed in soft, warm interior light, creating subtle highlights on the car's plastic shell and the polished floor, establishing a clean, focused, and tranquil atmosphere that isolates the simple, elegant physics of motion.", + "seed": 5196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_1/index_9846_seed_5196_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0069795969873666765, + "vlm_logprob": -0.0069795969873666765, + "vlm_yes_no_margin": 5.675, + "vlm_token_credit": -0.0069795969873666765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004428816959261894 + ], + "opposite_logprob_tokens": [ + -6.254428863525391 + ], + "logprob_reward": -0.004428816959261894, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.004428816959261894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07683534920215607, + 0.04800352826714516, + 0.03273169696331024, + 0.1103823184967041, + 0.04402374476194382, + 0.08662279695272446, + 0.04891159385442734, + 0.37927523255348206, + 0.11496280878782272, + 0.05825097858905792 + ], + "credit_map": [ + 0.007225510198622942, + 0.011428294703364372, + 0.013629617169499397, + 0.009656221605837345, + 0.01552213542163372, + 0.0047090849839150906, + 0.010035949759185314, + 0.004628530703485012, + 0.002745042322203517, + 0.010368384420871735, + 0.006149117834866047, + 0.01422430295497179, + 0.005421757698059082, + 0.003434586338698864, + 0.0037853028625249863, + 0.0018750332528725266, + 0.0038191648200154305, + 0.0036763441748917103, + 0.0032545472495257854, + 0.002963473554700613, + 0.006703014485538006, + 0.006026912946254015, + 0.0036081408616155386, + 0.0026801000349223614, + 0.01185702160000801, + 0.009924516081809998, + 0.013709046877920628, + 0.01812722161412239, + 0.00957540050148964, + 0.01203849446028471, + 0.025090061128139496, + 0.010060562752187252, + 0.004533193539828062, + 0.0035700956359505653, + 0.010834535583853722, + 0.0038754481356590986, + 0.005353706888854504, + 0.006731322966516018, + 0.004189879167824984, + 0.004935559816658497, + 0.011201173067092896, + 0.008947517722845078, + 0.010661179199814796, + 0.011912701651453972, + 0.012614498846232891, + 0.006427693646401167, + 0.01325710117816925, + 0.011600937694311142, + 0.002686307765543461, + 0.007650513201951981, + 0.005914729088544846, + 0.009430506266653538, + 0.008032099343836308, + 0.0056118774227797985, + 0.0060140290297567844, + 0.0035715284757316113, + 0.03684753179550171, + 0.02836361899971962, + 0.08151399344205856, + 0.03874130919575691, + 0.03123035468161106, + 0.0651945248246193, + 0.027403876185417175, + 0.06998002529144287, + 0.01239586528390646, + 0.01281858328729868, + 0.009063205681741238, + 0.01280997134745121, + 0.024068810045719147, + 0.011243166401982307, + 0.019083332270383835, + 0.013479867950081825, + 0.005318829324096441, + 0.004549507051706314, + 0.009889455512166023, + 0.006559624336659908, + 0.004742741584777832, + 0.008478441275656223, + 0.007724577095359564, + 0.010987803339958191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.004428816959261894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028269824106246233 + ], + "opposite_logprob_tokens": [ + -5.877827167510986 + ], + "logprob_reward": -0.0028269824106246233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028269824106246233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08409085124731064, + 0.06878509372472763, + 0.17864938080310822, + 0.060839857906103134, + 0.041741520166397095, + 0.08555731922388077, + 0.04812939837574959, + 0.32563701272010803, + 0.0691058337688446, + 0.03746373951435089 + ], + "credit_map": [ + 0.010126624256372452, + 0.012584765441715717, + 0.012248770333826542, + 0.012802415527403355, + 0.017317678779363632, + 0.005212110932916403, + 0.008634409867227077, + 0.005164082627743483, + 0.006701528560370207, + 0.010510744526982307, + 0.011583969928324223, + 0.013945616781711578, + 0.0075722914189100266, + 0.006261793896555901, + 0.007965265773236752, + 0.004243886098265648, + 0.020819129422307014, + 0.021663848310709, + 0.01836143434047699, + 0.010919895023107529, + 0.04772206395864487, + 0.030014079064130783, + 0.016524719074368477, + 0.012624210678040981, + 0.005272771697491407, + 0.005187937058508396, + 0.007816901430487633, + 0.010411333292722702, + 0.00603193836286664, + 0.007811057846993208, + 0.012492842972278595, + 0.005815073847770691, + 0.006021317094564438, + 0.005164844915270805, + 0.007326906081289053, + 0.0035076136700809, + 0.005866592284291983, + 0.005397220607846975, + 0.004019227810204029, + 0.004437795374542475, + 0.010916056111454964, + 0.007031526416540146, + 0.011499661952257156, + 0.013761023990809917, + 0.012920194305479527, + 0.007361710071563721, + 0.011335908435285091, + 0.010731233283877373, + 0.0031018159352242947, + 0.009753035381436348, + 0.006636068224906921, + 0.006051318719983101, + 0.005577778443694115, + 0.00635176757350564, + 0.0057771350257098675, + 0.004880479071289301, + 0.026763074100017548, + 0.01909698359668255, + 0.09046288579702377, + 0.024555357173085213, + 0.020104527473449707, + 0.04485642537474632, + 0.019186876714229584, + 0.08061087131500244, + 0.007465394213795662, + 0.007875396870076656, + 0.007387008052319288, + 0.007431849371641874, + 0.010024072602391243, + 0.009085644967854023, + 0.010281584225594997, + 0.009554876014590263, + 0.0030195601284503937, + 0.0024917463306337595, + 0.005786945577710867, + 0.005110766272991896, + 0.004383880645036697, + 0.0038547138683497906, + 0.006010814569890499, + 0.006805312354117632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.0028269824106246233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327572926878929 + ], + "opposite_logprob_tokens": [ + -3.77327561378479 + ], + "logprob_reward": -0.02327572926878929, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327572926878929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09715393930673599, + 0.11528688669204712, + 0.11902017146348953, + 0.06564920395612717, + 0.12321330606937408, + 0.30803999304771423, + 0.0182713083922863, + 0.06044667586684227, + 0.028016550466418266, + 0.0649019330739975 + ], + "credit_map": [ + 0.008787289261817932, + 0.010786964558064938, + 0.012451357208192348, + 0.02069532684981823, + 0.014555813744664192, + 0.008848226629197598, + 0.014755284413695335, + 0.006273679435253143, + 0.0068726870231330395, + 0.027810705825686455, + 0.012484347447752953, + 0.041041918098926544, + 0.008125358261168003, + 0.006504111457616091, + 0.009340026415884495, + 0.003107733093202114, + 0.01471671462059021, + 0.011427626013755798, + 0.01182583998888731, + 0.009035004302859306, + 0.02799662947654724, + 0.019870730116963387, + 0.012125018984079361, + 0.012022615410387516, + 0.006449402775615454, + 0.007007247768342495, + 0.007197200786322355, + 0.013911097310483456, + 0.0064958687871694565, + 0.007612497080117464, + 0.01062692329287529, + 0.0063489689491689205, + 0.008781587705016136, + 0.008973629213869572, + 0.025804979726672173, + 0.013303176499903202, + 0.019009893760085106, + 0.014571920037269592, + 0.018367011100053787, + 0.014401096850633621, + 0.026141438633203506, + 0.03248125687241554, + 0.03919116035103798, + 0.047871991991996765, + 0.056290995329618454, + 0.025952549651265144, + 0.04489142820239067, + 0.03521917015314102, + 0.0017758224857971072, + 0.002315758727490902, + 0.0025978023186326027, + 0.0017576212994754314, + 0.002733095781877637, + 0.002949478104710579, + 0.001851764041930437, + 0.0022899657487869263, + 0.005936077795922756, + 0.004677732940763235, + 0.011612283065915108, + 0.006104787811636925, + 0.005364608019590378, + 0.010708973743021488, + 0.0045930021442472935, + 0.011449212208390236, + 0.0030612479895353317, + 0.0030408999882638454, + 0.002676591509953141, + 0.003339701797813177, + 0.004222001414746046, + 0.0029123814310878515, + 0.0041301376186311245, + 0.004633587319403887, + 0.007294630166143179, + 0.005944469012320042, + 0.012108800932765007, + 0.00975015014410019, + 0.0043823751620948315, + 0.008075174875557423, + 0.00675701629370451, + 0.010589312762022018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.02327572926878929 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031885281205177307 + ], + "opposite_logprob_tokens": [ + -5.753188610076904 + ], + "logprob_reward": -0.0031885281205177307, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031885281205177307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07053861767053604, + 0.05065946280956268, + 0.2564797103404999, + 0.052776046097278595, + 0.06488790363073349, + 0.18224729597568512, + 0.06732174754142761, + 0.05627690628170967, + 0.07028120011091232, + 0.12853103876113892 + ], + "credit_map": [ + 0.007799503859132528, + 0.01039110217243433, + 0.011102722026407719, + 0.009109498001635075, + 0.01193248201161623, + 0.005651501938700676, + 0.008783005177974701, + 0.005768807139247656, + 0.003574361326172948, + 0.007935441099107265, + 0.007119753398001194, + 0.013340314850211143, + 0.006589770317077637, + 0.0035113145131617785, + 0.006176832597702742, + 0.002411677036434412, + 0.026943815872073174, + 0.027033012360334396, + 0.02357415109872818, + 0.017096664756536484, + 0.06768403947353363, + 0.048371847718954086, + 0.021318094804883003, + 0.02445809170603752, + 0.004051239229738712, + 0.005992097780108452, + 0.008784766308963299, + 0.005701415706425905, + 0.004738327115774155, + 0.006787245161831379, + 0.010371065698564053, + 0.006349893286824226, + 0.00647561764344573, + 0.004825511947274208, + 0.013348638080060482, + 0.006919979117810726, + 0.008676741272211075, + 0.00873697828501463, + 0.007819276303052902, + 0.008085167966783047, + 0.021373840048909187, + 0.019151968881487846, + 0.023231780156493187, + 0.01858377270400524, + 0.03384426608681679, + 0.016619741916656494, + 0.02733648754656315, + 0.02210545912384987, + 0.004957285709679127, + 0.014916644431650639, + 0.012320983223617077, + 0.0056608859449625015, + 0.007014684844762087, + 0.009728336706757545, + 0.006444923114031553, + 0.006278005428612232, + 0.004745573736727238, + 0.0039887563325464725, + 0.01830017752945423, + 0.0036643361672759056, + 0.0033634882420301437, + 0.005011367611587048, + 0.002866307971999049, + 0.014336901716887951, + 0.0076105608604848385, + 0.006459493190050125, + 0.005867375992238522, + 0.005711423233151436, + 0.004917474929243326, + 0.012384869158267975, + 0.0048042056150734425, + 0.02252579852938652, + 0.011948688887059689, + 0.009156505577266216, + 0.020229043439030647, + 0.01589992642402649, + 0.007597483228892088, + 0.015470127575099468, + 0.02486858144402504, + 0.02336070127785206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 9 + ], + "selected_reward": -0.0031885281205177307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001177928177639842 + ], + "opposite_logprob_tokens": [ + -6.751177787780762 + ], + "logprob_reward": -0.001177928177639842, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001177928177639842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15278056263923645, + 0.04698879271745682, + 0.12437161058187485, + 0.11214280128479004, + 0.14725609123706818, + 0.18300551176071167, + 0.03967152163386345, + 0.03934913128614426, + 0.08061902225017548, + 0.07381495088338852 + ], + "credit_map": [ + 0.015444491058588028, + 0.02238163910806179, + 0.018482927232980728, + 0.01819949969649315, + 0.033554449677467346, + 0.01596911996603012, + 0.017127348110079765, + 0.011621102690696716, + 0.0035051906015723944, + 0.006338458973914385, + 0.008373056538403034, + 0.00899215042591095, + 0.00811130553483963, + 0.003433270612731576, + 0.005786559544503689, + 0.0024487979244440794, + 0.013865657150745392, + 0.013000701554119587, + 0.013940169475972652, + 0.010430377908051014, + 0.024195749312639236, + 0.020224938169121742, + 0.015638621523976326, + 0.013075391761958599, + 0.008568805642426014, + 0.009451810270547867, + 0.015507229603827, + 0.018239401280879974, + 0.012193064205348492, + 0.010716988705098629, + 0.026152752339839935, + 0.011312754824757576, + 0.014699545688927174, + 0.016050241887569427, + 0.0318305604159832, + 0.013083932921290398, + 0.01741655357182026, + 0.019017884507775307, + 0.018204985186457634, + 0.0169523935765028, + 0.01923886500298977, + 0.02067248523235321, + 0.02425171062350273, + 0.02563139609992504, + 0.02884758822619915, + 0.015813196077942848, + 0.02357499487698078, + 0.024975279346108437, + 0.0034637008793652058, + 0.005548426881432533, + 0.005412450060248375, + 0.004884833004325628, + 0.004389568697661161, + 0.005322716664522886, + 0.006053971126675606, + 0.004595853388309479, + 0.003625044133514166, + 0.005279141943901777, + 0.006398824043571949, + 0.004907103255391121, + 0.004852804355323315, + 0.004741167649626732, + 0.003883225377649069, + 0.0056618209928274155, + 0.01087066438049078, + 0.008657203055918217, + 0.01597176492214203, + 0.006392780691385269, + 0.010641814209520817, + 0.007413146086037159, + 0.007968620397150517, + 0.01270302664488554, + 0.007241153158247471, + 0.005351684521883726, + 0.01370300818234682, + 0.011811638250946999, + 0.006165610626339912, + 0.013146346434950829, + 0.008671368472278118, + 0.007724144030362368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.001177928177639842 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle close-up begins at the top of an inclined plane, as the camera begins a smooth, continuous downward tilt, carefully following the motion of a small toy car. The subject is a sleek, vibrant red plastic toy car, its simple molded body reflecting the ambient light, resting on an inclined unfinished pine plank, its grain clearly visible. The car is released from complete rest and begins its descent; its black plastic wheels begin to spin, accelerating rapidly and decisively across the wooden surface under the pull of gravity, quickly blurring into smooth motion. The camera maintains close tracking, capturing the seamless transition as the car rolls from the end of the ramp onto the polished light brown hardwood floor. Its momentum carries it forward across the flat surface, where it gradually slows, gliding silently for a few feet before coming to a gentle, complete stop. The entire scene is bathed in soft, warm interior light, creating subtle highlights on the car's plastic shell and the polished floor, establishing a clean, focused, and tranquil atmosphere that isolates the simple, elegant physics of motion.", + "seed": 5197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_1/index_9846_seed_5197_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02488389452919364, + "vlm_logprob": -0.02488389452919364, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.02488389452919364, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004371847957372665 + ], + "opposite_logprob_tokens": [ + -6.754371643066406 + ], + "logprob_reward": -0.004371847957372665, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.004371847957372665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14789143204689026, + 0.40724045038223267, + 0.03475496172904968, + 0.047228869050741196, + 0.06125030294060707, + 0.03901461884379387, + 0.05681479349732399, + 0.0493801087141037, + 0.07853055000305176, + 0.07789390534162521 + ], + "credit_map": [ + 0.022705717012286186, + 0.011192968115210533, + 0.023021426051855087, + 0.01315185148268938, + 0.010961548425257206, + 0.010693371295928955, + 0.03195340931415558, + 0.02421114221215248, + 0.024385951459407806, + 0.007413080893456936, + 0.02290794812142849, + 0.007059915456920862, + 0.012582836672663689, + 0.15332205593585968, + 0.012531779706478119, + 0.16703687608242035, + 0.0046930257230997086, + 0.004090769682079554, + 0.004911696072667837, + 0.005062973592430353, + 0.005824822932481766, + 0.004051447845995426, + 0.0035220603458583355, + 0.0025981643702834845, + 0.006772676482796669, + 0.004629576578736305, + 0.006635218858718872, + 0.006721222773194313, + 0.0036116193514317274, + 0.0059959786012768745, + 0.00471794418990612, + 0.008144634775817394, + 0.006438089068979025, + 0.005382683593779802, + 0.007842709310352802, + 0.008352904580533504, + 0.005795315373688936, + 0.011784967966377735, + 0.007603291422128677, + 0.008050341159105301, + 0.005295425187796354, + 0.003118600696325302, + 0.0063269902020692825, + 0.005017591640353203, + 0.004212080035358667, + 0.004822710528969765, + 0.0049294764176011086, + 0.005291744600981474, + 0.006976925302296877, + 0.0067321546375751495, + 0.01131721492856741, + 0.00554997194558382, + 0.0065588075667619705, + 0.007696177344769239, + 0.005635841283947229, + 0.0063477009534835815, + 0.00763643579557538, + 0.005331526976078749, + 0.004895289894193411, + 0.009636569768190384, + 0.004988864995539188, + 0.00552565511316061, + 0.005902164150029421, + 0.005463598296046257, + 0.023058611899614334, + 0.009386115707457066, + 0.010187448933720589, + 0.00843513011932373, + 0.005767162423580885, + 0.006845287978649139, + 0.005649187136441469, + 0.009201601147651672, + 0.013546469621360302, + 0.00890770461410284, + 0.011198623105883598, + 0.006984251085668802, + 0.007355116307735443, + 0.008142470382153988, + 0.00677890470251441, + 0.014980367384850979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.004371847957372665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008653167635202408 + ], + "opposite_logprob_tokens": [ + -4.758653163909912 + ], + "logprob_reward": -0.008653167635202408, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008653167635202408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12950584292411804, + 0.3916315734386444, + 0.11536311358213425, + 0.08531565964221954, + 0.07765641063451767, + 0.021062597632408142, + 0.03492758795619011, + 0.04851594567298889, + 0.06922390311956406, + 0.02679731510579586 + ], + "credit_map": [ + 0.020554902032017708, + 0.010276312939822674, + 0.022336026653647423, + 0.013848964124917984, + 0.00808427669107914, + 0.011438368819653988, + 0.0185384564101696, + 0.024428531527519226, + 0.029477743431925774, + 0.010164783336222172, + 0.026163047179579735, + 0.00966530293226242, + 0.011766239069402218, + 0.13709478080272675, + 0.01417625043541193, + 0.15312343835830688, + 0.019178465008735657, + 0.0096574192866683, + 0.022197425365447998, + 0.01190539076924324, + 0.010673326440155506, + 0.015034670941531658, + 0.010728721506893635, + 0.015987688675522804, + 0.016551921144127846, + 0.00938696414232254, + 0.012634072452783585, + 0.008487341925501823, + 0.006240080576390028, + 0.011721712537109852, + 0.007026353385299444, + 0.013267219066619873, + 0.009369902312755585, + 0.005438018590211868, + 0.005701267160475254, + 0.0054069627076387405, + 0.007567636668682098, + 0.022274412214756012, + 0.008041826076805592, + 0.013856383971869946, + 0.0033706529065966606, + 0.0015462235314771533, + 0.0030274305026978254, + 0.0024195476435124874, + 0.002190811326727271, + 0.0029215686954557896, + 0.001960556488484144, + 0.0036258064210414886, + 0.007495406549423933, + 0.0031998285558074713, + 0.007919247262179852, + 0.0035878189373761415, + 0.0023737188894301653, + 0.0031536289025098085, + 0.0025479206815361977, + 0.004650020971894264, + 0.010258183814585209, + 0.007245172746479511, + 0.004757843911647797, + 0.007646446116268635, + 0.0036356085911393166, + 0.004239273723214865, + 0.00453550461679697, + 0.006197912152856588, + 0.010391793213784695, + 0.009004741907119751, + 0.006853320170193911, + 0.008954138495028019, + 0.007003219798207283, + 0.008355227299034595, + 0.006579555571079254, + 0.012081905268132687, + 0.0035116213839501143, + 0.0031768225599080324, + 0.00410987576469779, + 0.003013096982613206, + 0.0032176547683775425, + 0.002641724655404687, + 0.00292291771620512, + 0.004203603602945805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.008653167635202408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028552263975143 + ], + "opposite_logprob_tokens": [ + -2.350285530090332 + ], + "logprob_reward": -0.10028552263975143, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028552263975143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17965419590473175, + 0.06637787818908691, + 0.09716398268938065, + 0.265439510345459, + 0.07187245041131973, + 0.07708943635225296, + 0.15108537673950195, + 0.02414887398481369, + 0.020228397101163864, + 0.04693993180990219 + ], + "credit_map": [ + 0.02781025692820549, + 0.014113045297563076, + 0.0440312959253788, + 0.02107921615242958, + 0.014170621521770954, + 0.009795435704290867, + 0.03723876178264618, + 0.011415560729801655, + 0.009454232640564442, + 0.007816364988684654, + 0.010075131431221962, + 0.005261349491775036, + 0.011667930521070957, + 0.009357539005577564, + 0.005742321722209454, + 0.007003006525337696, + 0.019392605870962143, + 0.008007543161511421, + 0.01615861803293228, + 0.012724698521196842, + 0.010443015955388546, + 0.011671265587210655, + 0.007949612103402615, + 0.01081661693751812, + 0.057776909321546555, + 0.0355854369699955, + 0.04163346067070961, + 0.025483231991529465, + 0.017730019986629486, + 0.03668268024921417, + 0.0204569473862648, + 0.030090810731053352, + 0.007887520827353, + 0.005189963150769472, + 0.005221405997872353, + 0.005806801840662956, + 0.00968756340444088, + 0.01784322038292885, + 0.008139519020915031, + 0.01209645252674818, + 0.00829397514462471, + 0.005798939149826765, + 0.012881999835371971, + 0.011219664476811886, + 0.008930052630603313, + 0.00969646219164133, + 0.010034045204520226, + 0.0102342888712883, + 0.02248355932533741, + 0.019979583099484444, + 0.03690972179174423, + 0.02023073472082615, + 0.009789993055164814, + 0.013437691144645214, + 0.012738295830786228, + 0.015515795908868313, + 0.0031927379313856363, + 0.0028762519359588623, + 0.002280108630657196, + 0.004543590359389782, + 0.00299260881729424, + 0.002780078211799264, + 0.0028642495162785053, + 0.0026192471850663424, + 0.0044889445416629314, + 0.0025452300906181335, + 0.0022397732827812433, + 0.0023908892180770636, + 0.0019221535185351968, + 0.0022839978337287903, + 0.0016955516766756773, + 0.0026618572883307934, + 0.007158244028687477, + 0.0069331880658864975, + 0.007756772451102734, + 0.004798361100256443, + 0.003265290055423975, + 0.004048354923725128, + 0.0038475377950817347, + 0.009132184088230133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.10028552263975143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977444276213646 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.00977444276213646, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977444276213646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11267854273319244, + 0.2640037536621094, + 0.06559262424707413, + 0.15967896580696106, + 0.056537311524152756, + 0.04034439101815224, + 0.05485062301158905, + 0.060162127017974854, + 0.09152672439813614, + 0.09462488442659378 + ], + "credit_map": [ + 0.015105902217328548, + 0.009399987757205963, + 0.027953462675213814, + 0.013650281354784966, + 0.010653533041477203, + 0.00728449784219265, + 0.022740449756383896, + 0.005890432745218277, + 0.01551610603928566, + 0.0048704445362091064, + 0.014316855929791927, + 0.005517385900020599, + 0.008748183026909828, + 0.0964316576719284, + 0.00900979619473219, + 0.10959333926439285, + 0.011109035462141037, + 0.005537837743759155, + 0.01159593090415001, + 0.006715215742588043, + 0.005445085931569338, + 0.008430597372353077, + 0.006131836213171482, + 0.010627083480358124, + 0.037945471704006195, + 0.02058650553226471, + 0.02439611218869686, + 0.01509039755910635, + 0.01141806598752737, + 0.021393178030848503, + 0.011412156745791435, + 0.017437081784009933, + 0.005591259337961674, + 0.003518378362059593, + 0.004323986824601889, + 0.005115535110235214, + 0.006041208282113075, + 0.01711796224117279, + 0.004362167324870825, + 0.010466815903782845, + 0.005753349978476763, + 0.00325122126378119, + 0.006078967358916998, + 0.005657809320837259, + 0.003819816280156374, + 0.004562415648251772, + 0.004754932131618261, + 0.0064658778719604015, + 0.009635317139327526, + 0.00638768682256341, + 0.013567888177931309, + 0.005905871279537678, + 0.0037896307185292244, + 0.004691537469625473, + 0.00411343015730381, + 0.006759259384125471, + 0.0114518404006958, + 0.00704621896147728, + 0.005924979690462351, + 0.010117962025105953, + 0.006306102499365807, + 0.006357759237289429, + 0.0066966102458536625, + 0.006260653957724571, + 0.01817767694592476, + 0.012637936510145664, + 0.010153556242585182, + 0.011061464436352253, + 0.008858175948262215, + 0.010309130884706974, + 0.007030801381915808, + 0.013297987170517445, + 0.011254283599555492, + 0.014479683712124825, + 0.016140541061758995, + 0.0108947129920125, + 0.007200940977782011, + 0.009175216779112816, + 0.007939969189465046, + 0.01753954403102398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.00977444276213646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013344916515052319 + ], + "opposite_logprob_tokens": [ + -6.6263346672058105 + ], + "logprob_reward": -0.0013344916515052319, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013344916515052319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.164480522274971, + 0.14848902821540833, + 0.035155221819877625, + 0.27229490876197815, + 0.07530692219734192, + 0.07855614274740219, + 0.0980868861079216, + 0.04192368686199188, + 0.04013475030660629, + 0.045571886003017426 + ], + "credit_map": [ + 0.024916566908359528, + 0.011611984111368656, + 0.022702641785144806, + 0.02015741728246212, + 0.013074198737740517, + 0.01326752919703722, + 0.039294034242630005, + 0.0194561630487442, + 0.01000231597572565, + 0.00599490525200963, + 0.010689474642276764, + 0.003986069932579994, + 0.006718258373439312, + 0.04986480623483658, + 0.006086435168981552, + 0.055146779865026474, + 0.004926036112010479, + 0.003768134629353881, + 0.004399355035275221, + 0.004568169359117746, + 0.004770029801875353, + 0.0044861589558422565, + 0.003718732390552759, + 0.004518608562648296, + 0.0554550364613533, + 0.03496374562382698, + 0.04126090928912163, + 0.027221612632274628, + 0.01941017061471939, + 0.03836149349808693, + 0.020317643880844116, + 0.035304319113492966, + 0.010120514780282974, + 0.005858095362782478, + 0.006538028828799725, + 0.007855792529881, + 0.007434531580656767, + 0.01743432879447937, + 0.007848803885281086, + 0.012216834351420403, + 0.012321121990680695, + 0.005670101381838322, + 0.009461498819291592, + 0.01010100357234478, + 0.004923091270029545, + 0.010566379874944687, + 0.008840454742312431, + 0.01667250134050846, + 0.020064696669578552, + 0.011375503614544868, + 0.025318779051303864, + 0.011117382906377316, + 0.004962977021932602, + 0.008140736259520054, + 0.006868432741612196, + 0.01023839320987463, + 0.004448223859071732, + 0.004637577570974827, + 0.0038582871202379465, + 0.006937805097550154, + 0.005763841327279806, + 0.004540863912552595, + 0.005932327825576067, + 0.005804763175547123, + 0.006674256641417742, + 0.004618608858436346, + 0.004358038306236267, + 0.005411121528595686, + 0.004298291634768248, + 0.004093697294592857, + 0.0038424504455178976, + 0.006838287692517042, + 0.006334214471280575, + 0.00783402007073164, + 0.006498198490589857, + 0.004532230086624622, + 0.004981436301022768, + 0.003970425110310316, + 0.005501586478203535, + 0.0059197829104959965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0013344916515052319 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle close-up begins at the top of an inclined plane, as the camera begins a smooth, continuous downward tilt, carefully following the motion of a small toy car. The subject is a sleek, vibrant red plastic toy car, its simple molded body reflecting the ambient light, resting on an inclined unfinished pine plank, its grain clearly visible. The car is released from complete rest and begins its descent; its black plastic wheels begin to spin, accelerating rapidly and decisively across the wooden surface under the pull of gravity, quickly blurring into smooth motion. The camera maintains close tracking, capturing the seamless transition as the car rolls from the end of the ramp onto the polished light brown hardwood floor. Its momentum carries it forward across the flat surface, where it gradually slows, gliding silently for a few feet before coming to a gentle, complete stop. The entire scene is bathed in soft, warm interior light, creating subtle highlights on the car's plastic shell and the polished floor, establishing a clean, focused, and tranquil atmosphere that isolates the simple, elegant physics of motion.", + "seed": 5198, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_1/index_9846_seed_5198_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.009731311653740703, + "vlm_logprob": -0.009731311653740703, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.009731311653740703, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005274072289466858 + ], + "opposite_logprob_tokens": [ + -6.755274295806885 + ], + "logprob_reward": -0.005274072289466858, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.005274072289466858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11502740532159805, + 0.07234140485525131, + 0.10744256526231766, + 0.11460495740175247, + 0.08371647447347641, + 0.04481472820043564, + 0.29906216263771057, + 0.05272704362869263, + 0.0646103098988533, + 0.04565298929810524 + ], + "credit_map": [ + 0.012404609471559525, + 0.011373097077012062, + 0.02608960121870041, + 0.012496301904320717, + 0.011006985791027546, + 0.012363807298243046, + 0.018309075385332108, + 0.010983915999531746, + 0.008584803901612759, + 0.011300474405288696, + 0.009339109063148499, + 0.011099377647042274, + 0.009440661408007145, + 0.009132995270192623, + 0.0057070255279541016, + 0.0077369497157633305, + 0.007406572345644236, + 0.016914749518036842, + 0.01293716486543417, + 0.020724570378661156, + 0.012283596210181713, + 0.01284314226359129, + 0.011828897520899773, + 0.012503867037594318, + 0.013743069022893906, + 0.01736956648528576, + 0.014824051409959793, + 0.017776619642972946, + 0.017430035397410393, + 0.013532508164644241, + 0.009413749910891056, + 0.010515348985791206, + 0.011420759372413158, + 0.009281313978135586, + 0.009337223134934902, + 0.0059973979368805885, + 0.012099926359951496, + 0.009909214451909065, + 0.009388871490955353, + 0.01628175750374794, + 0.005385579075664282, + 0.005487041547894478, + 0.004105960950255394, + 0.004987645428627729, + 0.008280220441520214, + 0.0071992031298577785, + 0.0047501870431005955, + 0.004618888255208731, + 0.022523919120430946, + 0.043673910200595856, + 0.0299476720392704, + 0.031195811927318573, + 0.039877042174339294, + 0.03675546497106552, + 0.049628354609012604, + 0.045459944754838943, + 0.009192013181746006, + 0.008228744380176067, + 0.005954391323029995, + 0.006804443430155516, + 0.006066401023417711, + 0.005046526435762644, + 0.004742757882922888, + 0.006691764574497938, + 0.013087483122944832, + 0.012766576372087002, + 0.010058979503810406, + 0.006094370037317276, + 0.00674121780321002, + 0.005013017915189266, + 0.005215322133153677, + 0.00563333323225379, + 0.009228800423443317, + 0.00966102909296751, + 0.007364784833043814, + 0.005613775458186865, + 0.004119592253118753, + 0.004209195263683796, + 0.0024319954682141542, + 0.00302381394430995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.005274072289466858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647050518542528 + ], + "opposite_logprob_tokens": [ + -5.379647254943848 + ], + "logprob_reward": -0.004647050518542528, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647050518542528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1490936428308487, + 0.22041387856006622, + 0.05436297878623009, + 0.10980977863073349, + 0.061434999108314514, + 0.1387799084186554, + 0.1274024397134781, + 0.06236378103494644, + 0.0628855749964714, + 0.013453030027449131 + ], + "credit_map": [ + 0.015985555946826935, + 0.020787810906767845, + 0.03386755660176277, + 0.015119001269340515, + 0.015624061226844788, + 0.012170336209237576, + 0.016771504655480385, + 0.018767807632684708, + 0.030238429084420204, + 0.02596680074930191, + 0.01998155750334263, + 0.024937013164162636, + 0.029423074796795845, + 0.04398074373602867, + 0.021109670400619507, + 0.02477656863629818, + 0.004679054953157902, + 0.008144669234752655, + 0.007414211053401232, + 0.007781599182635546, + 0.004874856676906347, + 0.009892978705465794, + 0.0056007299572229385, + 0.005974875763058662, + 0.012207617051899433, + 0.021140368655323982, + 0.011534319259226322, + 0.01397575531154871, + 0.02783476933836937, + 0.009027526713907719, + 0.006864028982818127, + 0.00722538772970438, + 0.007882198318839073, + 0.005519307218492031, + 0.006297021172940731, + 0.005009825341403484, + 0.004438415169715881, + 0.005378353409469128, + 0.008256514556705952, + 0.01865336112678051, + 0.021197471767663956, + 0.02088019624352455, + 0.012554359622299671, + 0.01530968677252531, + 0.020549897104501724, + 0.019740838557481766, + 0.01336450781673193, + 0.015182942152023315, + 0.010609487071633339, + 0.01910051889717579, + 0.012871824204921722, + 0.01476010400801897, + 0.018479520455002785, + 0.015453699976205826, + 0.020863408222794533, + 0.015263868495821953, + 0.00968836061656475, + 0.009190882556140423, + 0.006362547632306814, + 0.00863683968782425, + 0.008133206516504288, + 0.00630376348271966, + 0.005339623894542456, + 0.008708552457392216, + 0.007100316230207682, + 0.012277101166546345, + 0.008941167034208775, + 0.006978526711463928, + 0.007407882250845432, + 0.005938703194260597, + 0.00740298954769969, + 0.006838884204626083, + 0.002145913429558277, + 0.0020699654705822468, + 0.0021074824035167694, + 0.001948503777384758, + 0.001507352921180427, + 0.0015843695728108287, + 0.0008233420085161924, + 0.0012660990469157696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.004647050518542528 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633504942059517 + ], + "opposite_logprob_tokens": [ + -3.6513350009918213 + ], + "logprob_reward": -0.02633504942059517, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633504942059517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15759454667568207, + 0.09813845157623291, + 0.07687686383724213, + 0.12519718706607819, + 0.08784081041812897, + 0.12296668440103531, + 0.19822169840335846, + 0.048282988369464874, + 0.06826064735651016, + 0.016620146110653877 + ], + "credit_map": [ + 0.014568732120096684, + 0.019569184631109238, + 0.019806021824479103, + 0.021053174510598183, + 0.02090510167181492, + 0.015420530922710896, + 0.028312498703598976, + 0.017959311604499817, + 0.012294833548367023, + 0.012766532599925995, + 0.017624523490667343, + 0.011972120963037014, + 0.01134549267590046, + 0.013299661688506603, + 0.00959643255919218, + 0.009238853119313717, + 0.005720947869122028, + 0.011545583605766296, + 0.00901114009320736, + 0.013702411204576492, + 0.009942063130438328, + 0.009575079195201397, + 0.008849195204675198, + 0.008530437014997005, + 0.012442880310118198, + 0.026019906625151634, + 0.01090456172823906, + 0.015745816752314568, + 0.033424604684114456, + 0.010730217210948467, + 0.00805667694658041, + 0.007872524671256542, + 0.02591003105044365, + 0.008093840442597866, + 0.01927977241575718, + 0.005419481545686722, + 0.008540735580027103, + 0.005581357516348362, + 0.006300177425146103, + 0.008715410716831684, + 0.016289308667182922, + 0.018336808308959007, + 0.01193162053823471, + 0.01593424566090107, + 0.015217963606119156, + 0.018121855333447456, + 0.014588579535484314, + 0.012546312063932419, + 0.014832347631454468, + 0.027504779398441315, + 0.019516879692673683, + 0.022405702620744705, + 0.027003489434719086, + 0.0247648935765028, + 0.033509112894535065, + 0.02868448756635189, + 0.006735133472830057, + 0.00847538374364376, + 0.005257854703813791, + 0.005992543883621693, + 0.005373465362936258, + 0.005078771151602268, + 0.004432407673448324, + 0.006937430240213871, + 0.011415015906095505, + 0.013131187297403812, + 0.009890732355415821, + 0.009120892733335495, + 0.0068610976450145245, + 0.005714850965887308, + 0.006241173017770052, + 0.005885697901248932, + 0.0031952273566275835, + 0.002753867767751217, + 0.0021783483680337667, + 0.0018544449703767896, + 0.0014951475895941257, + 0.0017697945004329085, + 0.0015757693909108639, + 0.001797545119188726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.02633504942059517 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066370643675327 + ], + "opposite_logprob_tokens": [ + -4.511066436767578 + ], + "logprob_reward": -0.011066370643675327, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066370643675327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15223868191242218, + 0.05706510692834854, + 0.06453189998865128, + 0.05806858837604523, + 0.0931987389922142, + 0.19946086406707764, + 0.047386687248945236, + 0.039163365960121155, + 0.20794451236724854, + 0.08094161748886108 + ], + "credit_map": [ + 0.014270284213125706, + 0.0189189612865448, + 0.03266362473368645, + 0.016182221472263336, + 0.019738826900720596, + 0.014175433665513992, + 0.019058706238865852, + 0.017230618745088577, + 0.008621076121926308, + 0.008791401982307434, + 0.006537980865687132, + 0.006228670012205839, + 0.00776949105784297, + 0.00800531916320324, + 0.00533163920044899, + 0.005779528524726629, + 0.00678174989297986, + 0.006155918352305889, + 0.012169839814305305, + 0.009619280695915222, + 0.005452604033052921, + 0.010999887250363827, + 0.006169620435684919, + 0.007182995788753033, + 0.010725531727075577, + 0.008691908791661263, + 0.010143643245100975, + 0.004837261512875557, + 0.009888711385428905, + 0.004802833776921034, + 0.0038803420029580593, + 0.00509835546836257, + 0.01249640341848135, + 0.010047140531241894, + 0.008379992097616196, + 0.006732520647346973, + 0.00974330399185419, + 0.009065564721822739, + 0.010816308669745922, + 0.025917500257492065, + 0.02506173960864544, + 0.029033996164798737, + 0.018308395519852638, + 0.01805119961500168, + 0.032464586198329926, + 0.03887351602315903, + 0.016597861424088478, + 0.021069558337330818, + 0.0052672941237688065, + 0.006798447575420141, + 0.004751494619995356, + 0.005653962958604097, + 0.007654039654880762, + 0.0059476373717188835, + 0.005498247686773539, + 0.005815561395138502, + 0.006347894202917814, + 0.006075052544474602, + 0.004781798459589481, + 0.004762816708534956, + 0.004819216672331095, + 0.005028651095926762, + 0.0028401832096278667, + 0.004507752601057291, + 0.04802120476961136, + 0.04188687726855278, + 0.03370426595211029, + 0.025304950773715973, + 0.019586119800806046, + 0.01433598157018423, + 0.013812759891152382, + 0.011292350478470325, + 0.01242308784276247, + 0.017343157902359962, + 0.011867407709360123, + 0.013516559265553951, + 0.0074238525703549385, + 0.007736541330814362, + 0.004337344318628311, + 0.006293671205639839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.011066370643675327 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, vibrant red plastic toy car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car start from rest on an inclined unfinished pine plank with visible grain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera begin with a high-angle close-up and execute a smooth downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the car roll from the plank onto a polished light brown hardwood floor and come to a complete stop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, warm interior light creating highlights on the car and floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013340153964236379 + ], + "opposite_logprob_tokens": [ + -6.626334190368652 + ], + "logprob_reward": -0.0013340153964236379, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013340153964236379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13181698322296143, + 0.0785946398973465, + 0.0783403217792511, + 0.12224442511796951, + 0.0709381103515625, + 0.0735412985086441, + 0.11271368712186813, + 0.08435007929801941, + 0.08094499260187149, + 0.16651542484760284 + ], + "credit_map": [ + 0.019706904888153076, + 0.015027590095996857, + 0.018989762291312218, + 0.01622869074344635, + 0.01610526815056801, + 0.014792758971452713, + 0.016968214884400368, + 0.013997804373502731, + 0.010673142038285732, + 0.011938528157770634, + 0.011327521875500679, + 0.008005389012396336, + 0.009847204200923443, + 0.010610288940370083, + 0.007454320788383484, + 0.008738241158425808, + 0.007765473332256079, + 0.007375402841717005, + 0.013086375780403614, + 0.01000479981303215, + 0.008633242920041084, + 0.010868914425373077, + 0.010159458965063095, + 0.010446647182106972, + 0.015505391173064709, + 0.025000831112265587, + 0.016000134870409966, + 0.019241219386458397, + 0.014474941417574883, + 0.011505750939249992, + 0.009639852680265903, + 0.010876313783228397, + 0.0110328895971179, + 0.00873923022300005, + 0.007813955657184124, + 0.004400980193167925, + 0.0118697015568614, + 0.008860157802700996, + 0.006403679959475994, + 0.01181752234697342, + 0.01102715265005827, + 0.009002942591905594, + 0.00767602538689971, + 0.007975352928042412, + 0.01342045422643423, + 0.011236803606152534, + 0.006467823870480061, + 0.006734744179993868, + 0.008973502553999424, + 0.014102722518146038, + 0.011066821403801441, + 0.013501102104783058, + 0.012392592616379261, + 0.011605114676058292, + 0.020756632089614868, + 0.020315201953053474, + 0.012021495960652828, + 0.01079383585602045, + 0.009240623563528061, + 0.008304116316139698, + 0.011335672810673714, + 0.010134649462997913, + 0.00946335680782795, + 0.013056328520178795, + 0.01804182305932045, + 0.012873219326138496, + 0.012551741674542427, + 0.009210275486111641, + 0.007672006264328957, + 0.006928205490112305, + 0.005975200328975916, + 0.007692523300647736, + 0.034213148057460785, + 0.03619929030537605, + 0.026631437242031097, + 0.022173745557665825, + 0.011491082608699799, + 0.012302539311349392, + 0.011892609298229218, + 0.011611570604145527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0013340153964236379 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_10/index_tensor([6015, 6015])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_10/index_tensor([6015, 6015])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..397b1f7bcd27bb82d6d1da6a61f0b3506d98bc0b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_10/index_tensor([6015, 6015])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, tilted close-up focuses squarely on a massive bullfrog. Its skin, moist and smooth, speckled with varying shades of dark green and brown, crouches on a dark, rain-washed rock. In the soft, diffuse light characteristic of a rainy day, the enormous, translucent white vocal sac in its jaw is inflated to its fullest, like a thin-skinned balloon inflated, its surface trembling slightly with the immense tension. Its enormous, bulging, dark-gold eyes stare impassively ahead. In the background, a gurgling stream ripples through its surface from the steady drizzle. A thick white mist fills the air, transforming the distant shore and mountains into blurred, inky silhouettes of varying shades, reminiscent of the brushstrokes of a traditional ink landscape painting. The main tones of the entire picture are an ink-and-wash color scheme consisting of black, white, and gray, embellished only by the dark green of the bullfrog's skin and a few clumps of wet emerald green reeds in the foreground. The overall atmosphere, beneath the tranquil poetry of ink painting, is undercurrent-driven by the tilted perspective and the subject's poised state, interweaving a primal vitality with an eerie sense of unease.", + "seed": 30654, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_10/index_6015_seed_30654_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.054838737845420836, + "vlm_logprob": -0.054838737845420836, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.054838737845420836, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014270750805735588 + ], + "opposite_logprob_tokens": [ + -4.514270782470703 + ], + "logprob_reward": -0.014270750805735588, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014270750805735588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3004692792892456, + 0.08977261930704117, + 0.2940434515476227, + 0.04507751390337944, + 0.02617405727505684, + 0.03394380584359169, + 0.04704470559954643, + 0.06229467689990997, + 0.05606749653816223, + 0.04511236399412155 + ], + "credit_map": [ + 0.04048425331711769, + 0.030952753499150276, + 0.03765718266367912, + 0.04682832211256027, + 0.028874961659312248, + 0.03635343164205551, + 0.038283709436655045, + 0.041034698486328125, + 0.010192172601819038, + 0.008977332152426243, + 0.011267777532339096, + 0.013955727219581604, + 0.009845450520515442, + 0.008834781125187874, + 0.014245965518057346, + 0.012453409843146801, + 0.03128451481461525, + 0.03540033474564552, + 0.034156087785959244, + 0.03756600245833397, + 0.03302071988582611, + 0.03797900676727295, + 0.027548382058739662, + 0.057088423520326614, + 0.004477488808333874, + 0.005003171972930431, + 0.006390341557562351, + 0.00661759776994586, + 0.007784218527376652, + 0.004418839700520039, + 0.00563178351148963, + 0.004754077643156052, + 0.0033366100396960974, + 0.0027900536078959703, + 0.002839414868503809, + 0.002562756882980466, + 0.003018225310370326, + 0.0030363155528903008, + 0.0021652558352798223, + 0.006425428204238415, + 0.0035602832213044167, + 0.002532064216211438, + 0.0033912518993020058, + 0.006018158048391342, + 0.002826476003974676, + 0.0037607743870466948, + 0.0036410000175237656, + 0.008213798515498638, + 0.004891401622444391, + 0.005427673924714327, + 0.004677263554185629, + 0.0082058971747756, + 0.006125893909484148, + 0.005031832959502935, + 0.004917959216982126, + 0.007766785100102425, + 0.00584475276991725, + 0.00791183952242136, + 0.011334951967000961, + 0.007538911886513233, + 0.008102656342089176, + 0.006102776154875755, + 0.008039558306336403, + 0.0074192313477396965, + 0.006032467819750309, + 0.005246325396001339, + 0.00747615285217762, + 0.010715123265981674, + 0.0049726152792572975, + 0.005745481699705124, + 0.006474354770034552, + 0.009404976852238178, + 0.00520175788551569, + 0.004468155559152365, + 0.007291065063327551, + 0.006469043903052807, + 0.00560543779283762, + 0.004601977299898863, + 0.005396424327045679, + 0.0060785021632909775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014270750805735588 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026362333446741104 + ], + "opposite_logprob_tokens": [ + -3.651362419128418 + ], + "logprob_reward": -0.026362333446741104, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026362333446741104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34042659401893616, + 0.06425704807043076, + 0.20578297972679138, + 0.07697535306215286, + 0.03141486644744873, + 0.05094379931688309, + 0.030169343575835228, + 0.07725579291582108, + 0.08106431365013123, + 0.04170987755060196 + ], + "credit_map": [ + 0.03580542653799057, + 0.04278915375471115, + 0.04002244025468826, + 0.047444719821214676, + 0.038305651396512985, + 0.036782436072826385, + 0.052855927497148514, + 0.04642084240913391, + 0.009121058508753777, + 0.006383922882378101, + 0.010146614164113998, + 0.007423332426697016, + 0.006977222394198179, + 0.008818849921226501, + 0.006645164452493191, + 0.008740879595279694, + 0.02224757708609104, + 0.02982882596552372, + 0.020278772339224815, + 0.02440709061920643, + 0.024572523310780525, + 0.023438788950443268, + 0.017798611894249916, + 0.04321080073714256, + 0.007925933226943016, + 0.007931627333164215, + 0.010845903307199478, + 0.011362746357917786, + 0.008992395363748074, + 0.010024162009358406, + 0.011790537275373936, + 0.008102049119770527, + 0.0033505503088235855, + 0.003356973407790065, + 0.0032133448403328657, + 0.004989180713891983, + 0.004247365053743124, + 0.0036245072260499, + 0.003248042892664671, + 0.005384902935475111, + 0.005850634537637234, + 0.003918464761227369, + 0.005701450631022453, + 0.007554668001830578, + 0.004864610265940428, + 0.006767252925783396, + 0.004620749037712812, + 0.011665970087051392, + 0.0040720864199101925, + 0.0027756490744650364, + 0.0028933549765497446, + 0.00509977200999856, + 0.003115129191428423, + 0.0033514918759465218, + 0.0028226496651768684, + 0.0060392115265131, + 0.007768400013446808, + 0.009790441021323204, + 0.01407365407794714, + 0.009879265911877155, + 0.009234892204403877, + 0.007769329007714987, + 0.009560837410390377, + 0.009178979322314262, + 0.008747725747525692, + 0.008081850595772266, + 0.015403070487082005, + 0.011748935095965862, + 0.008358669467270374, + 0.007971738465130329, + 0.008896538987755775, + 0.011855782940983772, + 0.005837241653352976, + 0.0038220458664000034, + 0.005306129809468985, + 0.00606174860149622, + 0.0035871462896466255, + 0.005107037723064423, + 0.005099811125546694, + 0.006888717412948608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.026362333446741104 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021998509764671326 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.0021998509764671326, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021998509764671326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16202017664909363, + 0.1629779189825058, + 0.21320870518684387, + 0.06852488219738007, + 0.08936035633087158, + 0.07282830774784088, + 0.03480541333556175, + 0.062117092311382294, + 0.06772369146347046, + 0.06643343716859818 + ], + "credit_map": [ + 0.018460089340806007, + 0.01514491718262434, + 0.01793002150952816, + 0.0248687993735075, + 0.014489470049738884, + 0.01978256180882454, + 0.01717008650302887, + 0.03417423367500305, + 0.01396673172712326, + 0.01440668199211359, + 0.022392263635993004, + 0.022266119718551636, + 0.020871225744485855, + 0.021194113418459892, + 0.024489283561706543, + 0.02339150570333004, + 0.020718708634376526, + 0.024906210601329803, + 0.02606702409684658, + 0.033045560121536255, + 0.02419220469892025, + 0.022641941905021667, + 0.023080846294760704, + 0.038556214421987534, + 0.006836255080997944, + 0.0092255137860775, + 0.008041150867938995, + 0.007623749785125256, + 0.010012870654463768, + 0.0086341742426157, + 0.011235292069613934, + 0.0069158729165792465, + 0.013391428627073765, + 0.005851204041391611, + 0.009914085268974304, + 0.010515800677239895, + 0.00702004786580801, + 0.01684531755745411, + 0.0071851639077067375, + 0.018637314438819885, + 0.0075806681998074055, + 0.0063492972403764725, + 0.009177617728710175, + 0.010957720689475536, + 0.007243180647492409, + 0.009366334415972233, + 0.007204115856438875, + 0.014949370175600052, + 0.004486209712922573, + 0.0034440276212990284, + 0.004332283046096563, + 0.004191037733107805, + 0.004236648324877024, + 0.003247360698878765, + 0.0039185574278235435, + 0.006949288304895163, + 0.0060362666845321655, + 0.007209220435470343, + 0.010245353914797306, + 0.00975704938173294, + 0.007049178704619408, + 0.00655420683324337, + 0.008188752457499504, + 0.007077066693454981, + 0.0076819113455712795, + 0.00817660242319107, + 0.012686830945312977, + 0.008592403493821621, + 0.008302520960569382, + 0.0059630307368934155, + 0.0069411033764481544, + 0.009379290975630283, + 0.007720566354691982, + 0.008395628072321415, + 0.010666788555681705, + 0.0086453752592206, + 0.007905612699687481, + 0.006228318903595209, + 0.008474950678646564, + 0.008396201767027378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0021998509764671326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029791494831442833 + ], + "opposite_logprob_tokens": [ + -3.5297915935516357 + ], + "logprob_reward": -0.029791494831442833, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029791494831442833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2694582939147949, + 0.07253846526145935, + 0.1648029237985611, + 0.1071452870965004, + 0.09223249554634094, + 0.08378181606531143, + 0.035183876752853394, + 0.07688683271408081, + 0.06629111617803574, + 0.03167892247438431 + ], + "credit_map": [ + 0.03845485672354698, + 0.02913481369614601, + 0.029490718618035316, + 0.02742804028093815, + 0.02699413150548935, + 0.038857415318489075, + 0.03083815425634384, + 0.04826018959283829, + 0.008276337757706642, + 0.007360010873526335, + 0.010251384228467941, + 0.009462831541895866, + 0.008669352158904076, + 0.009001538157463074, + 0.007304447703063488, + 0.01221256610006094, + 0.02230663411319256, + 0.02121305838227272, + 0.01719772443175316, + 0.019246093928813934, + 0.021336384117603302, + 0.024982396513223648, + 0.014745115302503109, + 0.023775532841682434, + 0.012263673357665539, + 0.013217711821198463, + 0.012383637949824333, + 0.014125608839094639, + 0.013731756247580051, + 0.013678074814379215, + 0.015247775241732597, + 0.01249705720692873, + 0.015422143042087555, + 0.005166101735085249, + 0.008372597396373749, + 0.007697708439081907, + 0.007817576639354229, + 0.017823534086346626, + 0.0063590966165065765, + 0.02357373759150505, + 0.009537606500089169, + 0.006546693854033947, + 0.009648705832660198, + 0.012587549164891243, + 0.007947471924126148, + 0.010359645821154118, + 0.008053655736148357, + 0.01910049095749855, + 0.005116536282002926, + 0.0030311904847621918, + 0.003507462330162525, + 0.00436718575656414, + 0.002803114941343665, + 0.003874345915392041, + 0.003167120972648263, + 0.009316922165453434, + 0.008128443732857704, + 0.009533637203276157, + 0.011854903772473335, + 0.008364876732230186, + 0.009449505247175694, + 0.007693352177739143, + 0.009770660661160946, + 0.012091458775103092, + 0.007320302538573742, + 0.005991464946419001, + 0.010975359939038754, + 0.010773904621601105, + 0.00625081779435277, + 0.00657880213111639, + 0.006964483764022589, + 0.011435984633862972, + 0.004658654797822237, + 0.0029626842588186264, + 0.0041248188354074955, + 0.004578323103487492, + 0.003942479845136404, + 0.003734173718839884, + 0.0031024767085909843, + 0.004575313534587622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.029791494831442833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156925916671753 + ], + "opposite_logprob_tokens": [ + -1.7015693187713623 + ], + "logprob_reward": -0.20156925916671753, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156925916671753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42963024973869324, + 0.07705827802419662, + 0.22416938841342926, + 0.07735265046358109, + 0.02332269214093685, + 0.025035396218299866, + 0.023121962323784828, + 0.024924667552113533, + 0.0342145636677742, + 0.061170175671577454 + ], + "credit_map": [ + 0.07488720118999481, + 0.02571740187704563, + 0.0327296182513237, + 0.038843728601932526, + 0.03204214200377464, + 0.12459024786949158, + 0.03598129376769066, + 0.06483859568834305, + 0.008441613055765629, + 0.010910033248364925, + 0.010295598767697811, + 0.011089424602687359, + 0.008856982924044132, + 0.008101374842226505, + 0.007918513379991055, + 0.011444740928709507, + 0.02192307449877262, + 0.03165055066347122, + 0.02892071194946766, + 0.032255951315164566, + 0.02605595625936985, + 0.026279442012310028, + 0.023635637015104294, + 0.03344806656241417, + 0.009646367281675339, + 0.00952165201306343, + 0.007640815805643797, + 0.008063835091888905, + 0.012265028432011604, + 0.010837309993803501, + 0.007292063906788826, + 0.01208558026701212, + 0.002987317042425275, + 0.0020261292811483145, + 0.002869244432076812, + 0.002713192952796817, + 0.0019689316395670176, + 0.0028584422543644905, + 0.0027558039873838425, + 0.005143630318343639, + 0.0036827519070357084, + 0.0019723742734640837, + 0.0027478409465402365, + 0.0022913068532943726, + 0.0023274037521332502, + 0.0037414394319057465, + 0.0021401916164904833, + 0.006132086738944054, + 0.002497814828529954, + 0.0024061212316155434, + 0.0028143736999481916, + 0.003344010328873992, + 0.002962145023047924, + 0.002159660682082176, + 0.002720093820244074, + 0.004217743873596191, + 0.0022000158205628395, + 0.0030211168341338634, + 0.002877363469451666, + 0.00420872587710619, + 0.0035387526731938124, + 0.0026091313920915127, + 0.002281270222738385, + 0.0041882931254804134, + 0.003404035000130534, + 0.003995354752987623, + 0.005107251927256584, + 0.004529915284365416, + 0.0046541630290448666, + 0.0031223755795508623, + 0.004971207585185766, + 0.004430259112268686, + 0.0070015969686210155, + 0.0066468375734984875, + 0.009558900259435177, + 0.008860169909894466, + 0.007317570969462395, + 0.0055028414353728294, + 0.007851428352296352, + 0.008430829271674156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.20156925916671753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, tilted close-up focuses squarely on a massive bullfrog. Its skin, moist and smooth, speckled with varying shades of dark green and brown, crouches on a dark, rain-washed rock. In the soft, diffuse light characteristic of a rainy day, the enormous, translucent white vocal sac in its jaw is inflated to its fullest, like a thin-skinned balloon inflated, its surface trembling slightly with the immense tension. Its enormous, bulging, dark-gold eyes stare impassively ahead. In the background, a gurgling stream ripples through its surface from the steady drizzle. A thick white mist fills the air, transforming the distant shore and mountains into blurred, inky silhouettes of varying shades, reminiscent of the brushstrokes of a traditional ink landscape painting. The main tones of the entire picture are an ink-and-wash color scheme consisting of black, white, and gray, embellished only by the dark green of the bullfrog's skin and a few clumps of wet emerald green reeds in the foreground. The overall atmosphere, beneath the tranquil poetry of ink painting, is undercurrent-driven by the tilted perspective and the subject's poised state, interweaving a primal vitality with an eerie sense of unease.", + "seed": 30655, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_10/index_6015_seed_30655_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08963861558586358, + "vlm_logprob": -0.08963861558586358, + "vlm_yes_no_margin": 3.349999988079071, + "vlm_token_credit": -0.08963861558586358, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030422957614064217 + ], + "opposite_logprob_tokens": [ + -3.6554229259490967 + ], + "logprob_reward": -0.030422957614064217, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030422957614064217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29083240032196045, + 0.19878599047660828, + 0.11506207287311554, + 0.08450595289468765, + 0.04594870284199715, + 0.05708871781826019, + 0.06428547948598862, + 0.06392406672239304, + 0.03611644729971886, + 0.04345020651817322 + ], + "credit_map": [ + 0.040261294692754745, + 0.045462094247341156, + 0.026456372812390327, + 0.05302296578884125, + 0.021358748897910118, + 0.038821808993816376, + 0.03730885684490204, + 0.028140263631939888, + 0.033967770636081696, + 0.033803217113018036, + 0.017208077013492584, + 0.024640139192342758, + 0.01542022917419672, + 0.02367919497191906, + 0.026860106736421585, + 0.023207230493426323, + 0.012227882631123066, + 0.017507357522845268, + 0.013836377300322056, + 0.024332160130143166, + 0.009125707671046257, + 0.013524641282856464, + 0.015966719016432762, + 0.008541226387023926, + 0.009632684290409088, + 0.01448321994394064, + 0.0062176622450351715, + 0.01342281885445118, + 0.00923765730112791, + 0.012448560446500778, + 0.01118279155343771, + 0.007880551740527153, + 0.006790990009903908, + 0.006267454009503126, + 0.0041627436876297, + 0.005566948093473911, + 0.004174486733973026, + 0.00675191730260849, + 0.0062111662700772285, + 0.006022996734827757, + 0.007231787778437138, + 0.011653414927423, + 0.004043725784868002, + 0.010413676500320435, + 0.005346575286239386, + 0.005326078273355961, + 0.007946527563035488, + 0.005126933101564646, + 0.009957616217434406, + 0.009230399504303932, + 0.00590100884437561, + 0.009479940868914127, + 0.0052679735235869884, + 0.009914122521877289, + 0.00840198528021574, + 0.006132430396974087, + 0.008155737072229385, + 0.008913218975067139, + 0.004226370714604855, + 0.011704888194799423, + 0.00736077968031168, + 0.007593240588903427, + 0.011192641220986843, + 0.004777194000780582, + 0.004876105580478907, + 0.006078809499740601, + 0.003351842286065221, + 0.00460895337164402, + 0.0040106321685016155, + 0.005030886735767126, + 0.004915350582450628, + 0.0032438677735626698, + 0.006910267286002636, + 0.006232847459614277, + 0.004341001156717539, + 0.005589059088379145, + 0.003862046869471669, + 0.008300429210066795, + 0.004447272513061762, + 0.0037672813050448895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.030422957614064217 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026374178007245064 + ], + "opposite_logprob_tokens": [ + -3.651374101638794 + ], + "logprob_reward": -0.026374178007245064, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026374178007245064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23322875797748566, + 0.17844164371490479, + 0.03263474628329277, + 0.044420305639505386, + 0.0697614997625351, + 0.03827205300331116, + 0.09904474020004272, + 0.18460418283939362, + 0.05682425945997238, + 0.0627678632736206 + ], + "credit_map": [ + 0.032165855169296265, + 0.04116986319422722, + 0.01963970437645912, + 0.026225896552205086, + 0.020469022914767265, + 0.036877404898405075, + 0.02411886304616928, + 0.03256214037537575, + 0.021856186911463737, + 0.03234538808465004, + 0.01577882468700409, + 0.02560465596616268, + 0.018035857006907463, + 0.02139202132821083, + 0.02334544062614441, + 0.020083246752619743, + 0.0035625756718218327, + 0.004661060404032469, + 0.0043962737545371056, + 0.0051590087823569775, + 0.0031588359270244837, + 0.003949951846152544, + 0.0037769589107483625, + 0.003970076330006123, + 0.010434516705572605, + 0.004681608639657497, + 0.003558930242434144, + 0.005681184586137533, + 0.0048496052622795105, + 0.007675513159483671, + 0.004643063060939312, + 0.002895876532420516, + 0.010539507493376732, + 0.0077145700342953205, + 0.007698125205934048, + 0.006665354128926992, + 0.009293356910347939, + 0.011295570060610771, + 0.007877402938902378, + 0.008677607402205467, + 0.004671365022659302, + 0.007227672263979912, + 0.003607103368267417, + 0.0066324747167527676, + 0.0033511619549244642, + 0.0034849834628403187, + 0.0053510526195168495, + 0.003946238197386265, + 0.018335038796067238, + 0.009607170708477497, + 0.007978313602507114, + 0.008676033467054367, + 0.008541743271052837, + 0.016355544328689575, + 0.012556277215480804, + 0.01699460856616497, + 0.029390698298811913, + 0.020020650699734688, + 0.015674103051424026, + 0.034934431314468384, + 0.01669713668525219, + 0.019686654210090637, + 0.03311093896627426, + 0.015089542604982853, + 0.007941230200231075, + 0.0070751565508544445, + 0.0066163791343569756, + 0.0067053562961518764, + 0.005561108235269785, + 0.008107340894639492, + 0.007532951887696981, + 0.007284731604158878, + 0.009662823751568794, + 0.009446554817259312, + 0.007516937796026468, + 0.009111323393881321, + 0.0048569063656032085, + 0.010852019302546978, + 0.0063443719409406185, + 0.004976919386535883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.026374178007245064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643728025257587 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.004643728025257587, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643728025257587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19310107827186584, + 0.11551017314195633, + 0.11399468779563904, + 0.14450106024742126, + 0.07433129847049713, + 0.06979502737522125, + 0.07199111580848694, + 0.12783506512641907, + 0.030985578894615173, + 0.05795498564839363 + ], + "credit_map": [ + 0.026609299704432487, + 0.025634605437517166, + 0.01663634181022644, + 0.023973625153303146, + 0.017570478841662407, + 0.03479282930493355, + 0.02413175068795681, + 0.023752136155962944, + 0.021913092583417892, + 0.014400064013898373, + 0.01181167084723711, + 0.016901064664125443, + 0.009597428143024445, + 0.016574017703533173, + 0.014533435925841331, + 0.009779401123523712, + 0.01615370810031891, + 0.015215756371617317, + 0.009791810996830463, + 0.018046319484710693, + 0.014402700588107109, + 0.014636432752013206, + 0.01468751672655344, + 0.011060434393584728, + 0.02960786409676075, + 0.01898956671357155, + 0.0107607152312994, + 0.021674098446965218, + 0.012401910498738289, + 0.02668842487037182, + 0.015191666781902313, + 0.009186811745166779, + 0.010888848453760147, + 0.007682594005018473, + 0.008733871392905712, + 0.006077153608202934, + 0.006469266023486853, + 0.011140788905322552, + 0.009854655712842941, + 0.013484122231602669, + 0.007838974706828594, + 0.010534731671214104, + 0.00666401069611311, + 0.01083741057664156, + 0.007926774211227894, + 0.0107705257833004, + 0.009012691676616669, + 0.0062099117785692215, + 0.014183116145431995, + 0.007696771528571844, + 0.0051292311400175095, + 0.005719214677810669, + 0.0062804510816931725, + 0.01248096115887165, + 0.008908524177968502, + 0.011592849157750607, + 0.01372312568128109, + 0.019333990290760994, + 0.009813657030463219, + 0.02399607002735138, + 0.016293484717607498, + 0.016124282032251358, + 0.019861411303281784, + 0.008689029142260551, + 0.004047228954732418, + 0.005540903657674789, + 0.0035000466741621494, + 0.0041174013167619705, + 0.0025593035388737917, + 0.003675802843645215, + 0.003865314181894064, + 0.003679575165733695, + 0.008920300751924515, + 0.00884207896888256, + 0.00641266955062747, + 0.007925738580524921, + 0.005240901838988066, + 0.010521695017814636, + 0.005701780319213867, + 0.004389826208353043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.004643728025257587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097646087408066 + ], + "opposite_logprob_tokens": [ + -3.288097620010376 + ], + "logprob_reward": -0.038097646087408066, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097646087408066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24617289006710052, + 0.19225306808948517, + 0.03772825747728348, + 0.07028768211603165, + 0.09700524806976318, + 0.036737460643053055, + 0.10491843521595001, + 0.07701127231121063, + 0.03562440350651741, + 0.10226128250360489 + ], + "credit_map": [ + 0.03654669597744942, + 0.02766690030694008, + 0.024463342502713203, + 0.02721395716071129, + 0.02629791758954525, + 0.051341462880373, + 0.025001009926199913, + 0.027641596272587776, + 0.02643895149230957, + 0.015379125252366066, + 0.025296969339251518, + 0.027891894802451134, + 0.019757533445954323, + 0.038348834961652756, + 0.020148176699876785, + 0.01899157464504242, + 0.005568841937929392, + 0.004891679156571627, + 0.006232834421098232, + 0.003704765113070607, + 0.003490092698484659, + 0.005414029583334923, + 0.00336173502728343, + 0.00506427651271224, + 0.017489135265350342, + 0.006296967156231403, + 0.0050882878713309765, + 0.010836046189069748, + 0.004580400418490171, + 0.0163410771638155, + 0.005389692261815071, + 0.004266072995960712, + 0.015545310452580452, + 0.009929988533258438, + 0.010749603621661663, + 0.008759738877415657, + 0.011280401609838009, + 0.016408639028668404, + 0.011811516247689724, + 0.012520051561295986, + 0.004063678439706564, + 0.007806670852005482, + 0.002928317291662097, + 0.0048779072239995, + 0.0034039891324937344, + 0.004948411136865616, + 0.004782752599567175, + 0.003925735130906105, + 0.02078287862241268, + 0.011233579367399216, + 0.008909733965992928, + 0.008910275995731354, + 0.008486210368573666, + 0.018106315284967422, + 0.012837544083595276, + 0.01565188355743885, + 0.009818365797400475, + 0.00970499962568283, + 0.0074100494384765625, + 0.01372815202921629, + 0.007330060005187988, + 0.0067778658121824265, + 0.015432104468345642, + 0.006809662561863661, + 0.004418161232024431, + 0.0061185769736766815, + 0.004192625638097525, + 0.005153280217200518, + 0.0029632498044520617, + 0.004005152266472578, + 0.004492516629397869, + 0.004280836321413517, + 0.01547183096408844, + 0.013446290045976639, + 0.01072587352246046, + 0.013626993633806705, + 0.009433991275727749, + 0.02087395079433918, + 0.010163666680455208, + 0.008518685586750507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.038097646087408066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.348654568195343 + ], + "opposite_logprob_tokens": [ + -1.2236545085906982 + ], + "logprob_reward": -0.348654568195343, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.348654568195343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22755306959152222, + 0.14244520664215088, + 0.042031895369291306, + 0.04052358120679855, + 0.1080920621752739, + 0.06260558217763901, + 0.027895988896489143, + 0.18998754024505615, + 0.03690091148018837, + 0.1219642385840416 + ], + "credit_map": [ + 0.032110754400491714, + 0.02687970921397209, + 0.03141653910279274, + 0.02590373158454895, + 0.02090093493461609, + 0.042012035846710205, + 0.023848338052630424, + 0.024481000378727913, + 0.016894884407520294, + 0.022928986698389053, + 0.020156536251306534, + 0.018579505383968353, + 0.011773036792874336, + 0.021608376875519753, + 0.016712665557861328, + 0.013791198842227459, + 0.00705206673592329, + 0.007190646603703499, + 0.003631947562098503, + 0.005921164061874151, + 0.004252207465469837, + 0.005302215926349163, + 0.00559824425727129, + 0.0030834029894322157, + 0.007121601607650518, + 0.00696106581017375, + 0.003715415019541979, + 0.005697635468095541, + 0.004171391483396292, + 0.00628412002697587, + 0.0036093145608901978, + 0.002963034901767969, + 0.01556339394301176, + 0.01479425746947527, + 0.009752280078828335, + 0.013491316698491573, + 0.014225726947188377, + 0.016592679545283318, + 0.013976759277284145, + 0.009695629589259624, + 0.007970260456204414, + 0.010795550420880318, + 0.005514319520443678, + 0.008306511677801609, + 0.006602381821721792, + 0.009798943996429443, + 0.007149473764002323, + 0.006468135863542557, + 0.004348930902779102, + 0.0053372555412352085, + 0.0025613002944737673, + 0.002988749649375677, + 0.0034683721605688334, + 0.003592225955799222, + 0.003217819146811962, + 0.002381333615630865, + 0.02347707934677601, + 0.021394258365035057, + 0.01494000293314457, + 0.03647754341363907, + 0.020399611443281174, + 0.017995070666074753, + 0.039166536182165146, + 0.016137419268488884, + 0.005142364650964737, + 0.00445272633805871, + 0.0061778114177286625, + 0.004746717866510153, + 0.003348941681906581, + 0.005431810859590769, + 0.003971719648689032, + 0.0036288159899413586, + 0.018298018723726273, + 0.016768479719758034, + 0.015367655083537102, + 0.016253510490059853, + 0.010165509767830372, + 0.023151975125074387, + 0.011854114010930061, + 0.010104970075190067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.348654568195343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, tilted close-up focuses squarely on a massive bullfrog. Its skin, moist and smooth, speckled with varying shades of dark green and brown, crouches on a dark, rain-washed rock. In the soft, diffuse light characteristic of a rainy day, the enormous, translucent white vocal sac in its jaw is inflated to its fullest, like a thin-skinned balloon inflated, its surface trembling slightly with the immense tension. Its enormous, bulging, dark-gold eyes stare impassively ahead. In the background, a gurgling stream ripples through its surface from the steady drizzle. A thick white mist fills the air, transforming the distant shore and mountains into blurred, inky silhouettes of varying shades, reminiscent of the brushstrokes of a traditional ink landscape painting. The main tones of the entire picture are an ink-and-wash color scheme consisting of black, white, and gray, embellished only by the dark green of the bullfrog's skin and a few clumps of wet emerald green reeds in the foreground. The overall atmosphere, beneath the tranquil poetry of ink painting, is undercurrent-driven by the tilted perspective and the subject's poised state, interweaving a primal vitality with an eerie sense of unease.", + "seed": 30656, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_10/index_6015_seed_30656_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11065951236523688, + "vlm_logprob": -0.11065951236523688, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.11065951236523688, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013277918100357056 + ], + "opposite_logprob_tokens": [ + -4.513278007507324 + ], + "logprob_reward": -0.013277918100357056, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013277918100357056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15357185900211334, + 0.11778746545314789, + 0.17067964375019073, + 0.036627255380153656, + 0.14958348870277405, + 0.11040305346250534, + 0.11087960004806519, + 0.03432566300034523, + 0.06173710525035858, + 0.054404862225055695 + ], + "credit_map": [ + 0.016545483842492104, + 0.01631918177008629, + 0.017512889578938484, + 0.022566387429833412, + 0.013009849935770035, + 0.022050045430660248, + 0.030459899455308914, + 0.015108122490346432, + 0.010224232450127602, + 0.011732715182006359, + 0.009594837203621864, + 0.015463289804756641, + 0.00975783821195364, + 0.012532544322311878, + 0.03426751866936684, + 0.014214486815035343, + 0.016444509848952293, + 0.01818223111331463, + 0.021219389513134956, + 0.0451238639652729, + 0.01884252205491066, + 0.020568497478961945, + 0.016712302342057228, + 0.013586326502263546, + 0.005980703979730606, + 0.00463105970993638, + 0.004794440232217312, + 0.003947690594941378, + 0.002624821849167347, + 0.0034217368811368942, + 0.006412116345018148, + 0.004814682994037867, + 0.026192884892225266, + 0.01553354226052761, + 0.015015905722975731, + 0.018384162336587906, + 0.009868237189948559, + 0.021238865330815315, + 0.03151803836226463, + 0.011831854470074177, + 0.013294902630150318, + 0.010799297131597996, + 0.010413195006549358, + 0.018579186871647835, + 0.008675807155668736, + 0.012666188180446625, + 0.027316687628626823, + 0.008657783269882202, + 0.008899820037186146, + 0.009888638742268085, + 0.019596057012677193, + 0.018709825351834297, + 0.011393142864108086, + 0.015359359793365002, + 0.01675294153392315, + 0.01027980912476778, + 0.0057058874517679214, + 0.004536573309451342, + 0.003826592583209276, + 0.004680920392274857, + 0.0025851891841739416, + 0.00308420741930604, + 0.006512846332043409, + 0.003393446560949087, + 0.010158496908843517, + 0.005925892852246761, + 0.004856084007769823, + 0.006126616150140762, + 0.003919876180589199, + 0.01814306154847145, + 0.005369713995605707, + 0.007237365003675222, + 0.009670351631939411, + 0.00512954406440258, + 0.0039229425601661205, + 0.004755013156682253, + 0.003448519855737686, + 0.01371692307293415, + 0.006851478014141321, + 0.006910089869052172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.013277918100357056 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033718500286340714 + ], + "opposite_logprob_tokens": [ + -3.4087185859680176 + ], + "logprob_reward": -0.033718500286340714, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033718500286340714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2178928554058075, + 0.08311337232589722, + 0.09370888769626617, + 0.06366036087274551, + 0.18952828645706177, + 0.07645829021930695, + 0.039648570120334625, + 0.05239536240696907, + 0.12152400612831116, + 0.06207004189491272 + ], + "credit_map": [ + 0.022245783358812332, + 0.02010728232562542, + 0.02655618079006672, + 0.033002808690071106, + 0.025646351277828217, + 0.027432017028331757, + 0.03597167134284973, + 0.026930784806609154, + 0.008861588314175606, + 0.008418186567723751, + 0.008419916033744812, + 0.009242908097803593, + 0.008659031242132187, + 0.0071221874095499516, + 0.023250717669725418, + 0.009138845838606358, + 0.011030049994587898, + 0.008691876195371151, + 0.011913090012967587, + 0.020774805918335915, + 0.010890183039009571, + 0.008959723636507988, + 0.01469030324369669, + 0.006758852396160364, + 0.00653960881754756, + 0.005897924304008484, + 0.007010608911514282, + 0.007281736936420202, + 0.0065299104899168015, + 0.006077744998037815, + 0.0159673523157835, + 0.008355464786291122, + 0.031152235344052315, + 0.016595473513007164, + 0.019898710772395134, + 0.021477116271853447, + 0.013068080879747868, + 0.02695591188967228, + 0.04417981952428818, + 0.01620093546807766, + 0.009217050857841969, + 0.00940030813217163, + 0.00872631836682558, + 0.012413280084729195, + 0.006934614852070808, + 0.010112667456269264, + 0.013320759870111942, + 0.006333298049867153, + 0.0047061750665307045, + 0.003982465714216232, + 0.005888942163437605, + 0.005148767959326506, + 0.0035194121301174164, + 0.004673142917454243, + 0.0085710184648633, + 0.003158645471557975, + 0.006273876875638962, + 0.004983005579560995, + 0.004457694943994284, + 0.006248736288398504, + 0.004350557457655668, + 0.005415234714746475, + 0.015332898125052452, + 0.005333357956260443, + 0.021314967423677444, + 0.008915490470826626, + 0.007706892676651478, + 0.008023049682378769, + 0.0068685500882565975, + 0.04329073056578636, + 0.008307489566504955, + 0.01709682308137417, + 0.010542641393840313, + 0.006137945223599672, + 0.004958682227879763, + 0.004799309186637402, + 0.00355474348179996, + 0.018255366012454033, + 0.005405811592936516, + 0.008415540680289268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.033718500286340714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200207905843854 + ], + "opposite_logprob_tokens": [ + -6.127200126647949 + ], + "logprob_reward": -0.002200207905843854, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200207905843854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11902885884046555, + 0.1127016693353653, + 0.11603231728076935, + 0.04521642252802849, + 0.12045688182115555, + 0.06004239618778229, + 0.1110861524939537, + 0.03276862949132919, + 0.1591246873140335, + 0.12354198843240738 + ], + "credit_map": [ + 0.012329920195043087, + 0.01188567653298378, + 0.01532822847366333, + 0.015846990048885345, + 0.01340395025908947, + 0.01449308916926384, + 0.02234502322971821, + 0.013395977206528187, + 0.009289751760661602, + 0.009270450100302696, + 0.007629017811268568, + 0.010879913344979286, + 0.009089692495763302, + 0.012651903554797173, + 0.041112009435892105, + 0.012778935953974724, + 0.01203094981610775, + 0.013078993186354637, + 0.014723112806677818, + 0.030751368030905724, + 0.012243576347827911, + 0.013285108841955662, + 0.010982561856508255, + 0.008936643600463867, + 0.0054718623869121075, + 0.003843697253614664, + 0.003291197121143341, + 0.0040782783180475235, + 0.00457903603091836, + 0.004496729001402855, + 0.01180657185614109, + 0.00764905009418726, + 0.020498408004641533, + 0.010295276530086994, + 0.011785533279180527, + 0.013715110719203949, + 0.008346174843609333, + 0.01763213612139225, + 0.02724963240325451, + 0.0109346192330122, + 0.00779258506372571, + 0.006333916913717985, + 0.004280002322047949, + 0.007714088074862957, + 0.0053087580017745495, + 0.006864819675683975, + 0.01545534934848547, + 0.006292877718806267, + 0.010678363032639027, + 0.010551867075264454, + 0.022521618753671646, + 0.01850847154855728, + 0.011128803715109825, + 0.012704312801361084, + 0.014475465752184391, + 0.010517245158553123, + 0.00533886207267642, + 0.004739641677588224, + 0.0035682721063494682, + 0.0029581557027995586, + 0.0030349597800523043, + 0.005025619640946388, + 0.004195540212094784, + 0.003907579462975264, + 0.028879202902317047, + 0.01192894671112299, + 0.0096165481954813, + 0.010094386525452137, + 0.008578484877943993, + 0.05752735957503319, + 0.010373411700129509, + 0.022126348689198494, + 0.020427603274583817, + 0.01269253809005022, + 0.010375568643212318, + 0.009462827816605568, + 0.006479298695921898, + 0.036985982209444046, + 0.010786677710711956, + 0.016331495717167854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.002200207905843854 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789643362164497 + ], + "opposite_logprob_tokens": [ + -3.529789686203003 + ], + "logprob_reward": -0.029789643362164497, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789643362164497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20597343146800995, + 0.08726897835731506, + 0.1240222230553627, + 0.04224487766623497, + 0.1326061338186264, + 0.09420336037874222, + 0.09551655501127243, + 0.028225835412740707, + 0.15320274233818054, + 0.036735858768224716 + ], + "credit_map": [ + 0.019762197509407997, + 0.021324574947357178, + 0.021786607801914215, + 0.02971155196428299, + 0.025313688442111015, + 0.026514705270528793, + 0.03828819841146469, + 0.02327190525829792, + 0.010273204185068607, + 0.01135060004889965, + 0.006562719587236643, + 0.011408312246203423, + 0.0069947331212460995, + 0.01020041573792696, + 0.021243928000330925, + 0.009235058911144733, + 0.0115902628749609, + 0.012908303178846836, + 0.014503924176096916, + 0.029296020045876503, + 0.013126866891980171, + 0.014476717449724674, + 0.015081320889294147, + 0.013038812205195427, + 0.006045658607035875, + 0.0035326152574270964, + 0.00396343320608139, + 0.007574638817459345, + 0.0038337265141308308, + 0.004686100874096155, + 0.005760889034718275, + 0.006847816985100508, + 0.021251926198601723, + 0.011781143955886364, + 0.013691677711904049, + 0.016421301290392876, + 0.009226817637681961, + 0.01782333292067051, + 0.028832044452428818, + 0.013577883131802082, + 0.010905559174716473, + 0.009547630324959755, + 0.007767294999212027, + 0.012966600246727467, + 0.008753064088523388, + 0.01096259243786335, + 0.02403857558965683, + 0.009262039326131344, + 0.007614871487021446, + 0.008647837676107883, + 0.015936631709337234, + 0.01569679006934166, + 0.010622055269777775, + 0.01286647655069828, + 0.015828581526875496, + 0.008303312584757805, + 0.0033062731381505728, + 0.003323410637676716, + 0.0030533962417393923, + 0.0036282273940742016, + 0.0025061441119760275, + 0.0027618706226348877, + 0.0060341851785779, + 0.003612329950556159, + 0.024845903739333153, + 0.011231925338506699, + 0.008874217979609966, + 0.008611224591732025, + 0.008154504932463169, + 0.05620184540748596, + 0.009529444389045238, + 0.025753675028681755, + 0.005411638878285885, + 0.003657808294519782, + 0.004453198052942753, + 0.003698995104059577, + 0.002249605255201459, + 0.008191577158868313, + 0.00445361016318202, + 0.004619428887963295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.029789643362164497 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47431129217147827 + ], + "opposite_logprob_tokens": [ + -0.9743112921714783 + ], + "logprob_reward": -0.47431129217147827, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47431129217147827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16851797699928284, + 0.10621152818202972, + 0.270803302526474, + 0.0633421391248703, + 0.11093266308307648, + 0.08288070559501648, + 0.036414775997400284, + 0.06078101322054863, + 0.04045379161834717, + 0.05966215208172798 + ], + "credit_map": [ + 0.023754244670271873, + 0.024442218244075775, + 0.021156370639801025, + 0.017212824895977974, + 0.01686941646039486, + 0.02254306524991989, + 0.027721138671040535, + 0.014818700030446053, + 0.014610185287892818, + 0.016430793330073357, + 0.012421922758221626, + 0.018205352127552032, + 0.008126814849674702, + 0.010756940580904484, + 0.017781388014554977, + 0.007878139615058899, + 0.02906927838921547, + 0.027121176943182945, + 0.03352954238653183, + 0.07023816555738449, + 0.030930260196328163, + 0.030507715418934822, + 0.02555425465106964, + 0.02385292388498783, + 0.010670634917914867, + 0.009586703032255173, + 0.007561453152447939, + 0.007376091089099646, + 0.004760162904858589, + 0.0065902769565582275, + 0.012125218287110329, + 0.004671603441238403, + 0.016019629314541817, + 0.010202061384916306, + 0.01456944178789854, + 0.015710441395640373, + 0.008394477888941765, + 0.012780778110027313, + 0.02336219511926174, + 0.00989364180713892, + 0.010116656310856342, + 0.009757515043020248, + 0.009114321321249008, + 0.012935060076415539, + 0.008453462272882462, + 0.009417541325092316, + 0.014349771663546562, + 0.008736375719308853, + 0.0036215439904481173, + 0.0035541525576263666, + 0.004274897743016481, + 0.006022033281624317, + 0.0043990518897771835, + 0.0035249844659119844, + 0.008078347891569138, + 0.0029397618491202593, + 0.006579854991286993, + 0.0060192616656422615, + 0.008888430893421173, + 0.009495994076132774, + 0.005479084327816963, + 0.006839145440608263, + 0.010932601988315582, + 0.006546634249389172, + 0.005966144613921642, + 0.006751992739737034, + 0.0035840363707393408, + 0.0034447985235601664, + 0.0027418225072324276, + 0.0060523939318954945, + 0.004824467934668064, + 0.0070881349965929985, + 0.007864185608923435, + 0.008177182637155056, + 0.006452296860516071, + 0.0050522699020802975, + 0.005094893742352724, + 0.013323324732482433, + 0.007348694372922182, + 0.006349305156618357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.47431129217147827 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, tilted close-up focuses squarely on a massive bullfrog. Its skin, moist and smooth, speckled with varying shades of dark green and brown, crouches on a dark, rain-washed rock. In the soft, diffuse light characteristic of a rainy day, the enormous, translucent white vocal sac in its jaw is inflated to its fullest, like a thin-skinned balloon inflated, its surface trembling slightly with the immense tension. Its enormous, bulging, dark-gold eyes stare impassively ahead. In the background, a gurgling stream ripples through its surface from the steady drizzle. A thick white mist fills the air, transforming the distant shore and mountains into blurred, inky silhouettes of varying shades, reminiscent of the brushstrokes of a traditional ink landscape painting. The main tones of the entire picture are an ink-and-wash color scheme consisting of black, white, and gray, embellished only by the dark green of the bullfrog's skin and a few clumps of wet emerald green reeds in the foreground. The overall atmosphere, beneath the tranquil poetry of ink painting, is undercurrent-driven by the tilted perspective and the subject's poised state, interweaving a primal vitality with an eerie sense of unease.", + "seed": 30657, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_10/index_6015_seed_30657_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0949521193280816, + "vlm_logprob": -0.0949521193280816, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.0949521193280816, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012274795211851597 + ], + "opposite_logprob_tokens": [ + -4.637274742126465 + ], + "logprob_reward": -0.012274795211851597, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012274795211851597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21084244549274445, + 0.038177452981472015, + 0.023587286472320557, + 0.061338502913713455, + 0.0988268181681633, + 0.03636603802442551, + 0.05017954856157303, + 0.4248749911785126, + 0.033217933028936386, + 0.022588979452848434 + ], + "credit_map": [ + 0.06816428154706955, + 0.009759717620909214, + 0.022695237770676613, + 0.008057048544287682, + 0.006827662233263254, + 0.06867110729217529, + 0.01447420846670866, + 0.012193182483315468, + 0.005496303550899029, + 0.008016440086066723, + 0.006385415326803923, + 0.002914919750764966, + 0.0031655177008360624, + 0.0058024004101753235, + 0.0034912731498479843, + 0.002905181609094143, + 0.0047693257220089436, + 0.003079549642279744, + 0.0024500838480889797, + 0.0021189202088862658, + 0.002762960735708475, + 0.002413898939266801, + 0.003331582760438323, + 0.0026609646156430244, + 0.00699316943064332, + 0.0072951107285916805, + 0.015698598697781563, + 0.004619074985384941, + 0.0072489851154387, + 0.004271348938345909, + 0.00892200879752636, + 0.0062902094796299934, + 0.008762631565332413, + 0.00979122519493103, + 0.009607727639377117, + 0.013005862943828106, + 0.013256682083010674, + 0.009147897362709045, + 0.020669063553214073, + 0.014585732482373714, + 0.0068525043316185474, + 0.003912769723683596, + 0.006668944843113422, + 0.003946899902075529, + 0.0030844518914818764, + 0.004028039053082466, + 0.004815677180886269, + 0.00305675039999187, + 0.005810663569718599, + 0.005617421120405197, + 0.009348255582153797, + 0.004494300577789545, + 0.005815276876091957, + 0.006247438956052065, + 0.006830538157373667, + 0.00601565558463335, + 0.04273517057299614, + 0.022368049249053, + 0.11608850210905075, + 0.027511613443493843, + 0.02712978422641754, + 0.09883751720190048, + 0.0558820478618145, + 0.03432230278849602, + 0.002645101398229599, + 0.0029846394900232553, + 0.0036434258799999952, + 0.004982993006706238, + 0.0049074022099375725, + 0.002791822189465165, + 0.00633463729172945, + 0.004927909467369318, + 0.0019374870462343097, + 0.002288632560521364, + 0.00209063827060163, + 0.0031655514612793922, + 0.003401668043807149, + 0.0018538213334977627, + 0.004863479640334845, + 0.0029877021443098783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.012274795211851597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811738267540932 + ], + "opposite_logprob_tokens": [ + -3.2881174087524414 + ], + "logprob_reward": -0.03811738267540932, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811738267540932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3143244683742523, + 0.021718183532357216, + 0.03593732789158821, + 0.2962668538093567, + 0.09943579882383347, + 0.06268946081399918, + 0.026935512199997902, + 0.08912867307662964, + 0.028235608711838722, + 0.025328120216727257 + ], + "credit_map": [ + 0.10929907113313675, + 0.020900748670101166, + 0.016785256564617157, + 0.010838186368346214, + 0.011017859913408756, + 0.11547232419252396, + 0.01717350259423256, + 0.012837487272918224, + 0.0030927446205168962, + 0.005050952546298504, + 0.00273449276573956, + 0.001748644164763391, + 0.002505113836377859, + 0.0027216209564357996, + 0.002020871266722679, + 0.0018437413964420557, + 0.005099012516438961, + 0.004825534299015999, + 0.0047110458835959435, + 0.003942413255572319, + 0.004191325046122074, + 0.0029512967448681593, + 0.005753315053880215, + 0.004463382065296173, + 0.09411852061748505, + 0.01905136927962303, + 0.11280607432126999, + 0.014107576571404934, + 0.010951174423098564, + 0.015939190983772278, + 0.010215803980827332, + 0.019077118486166, + 0.01732250489294529, + 0.007273619994521141, + 0.013200120069086552, + 0.00968930497765541, + 0.009223412722349167, + 0.012278623878955841, + 0.018001990392804146, + 0.012446208856999874, + 0.011171897873282433, + 0.007883322425186634, + 0.013311783783137798, + 0.005927310790866613, + 0.005412377417087555, + 0.006528108846396208, + 0.0077169896103441715, + 0.0047376626171171665, + 0.004167256411164999, + 0.0029436820186674595, + 0.004227424971759319, + 0.0023927518632262945, + 0.003619037102907896, + 0.002445697085931897, + 0.004447543993592262, + 0.002692116890102625, + 0.00726740900427103, + 0.0024086241610348225, + 0.03419169411063194, + 0.003883316647261381, + 0.002599116414785385, + 0.02610601671040058, + 0.0038957148790359497, + 0.008776772767305374, + 0.0027404273860156536, + 0.0037784508895128965, + 0.003206556895747781, + 0.003949089907109737, + 0.004206061828881502, + 0.002114506671205163, + 0.00458714971318841, + 0.003653364023193717, + 0.004188400227576494, + 0.0035404388327151537, + 0.004870539531111717, + 0.0023217815905809402, + 0.0029221326112747192, + 0.00228718388825655, + 0.0029576404485851526, + 0.0022400026209652424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.03811738267540932 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616936504840851 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.003616936504840851, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616936504840851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2529616355895996, + 0.04440503567457199, + 0.0784953236579895, + 0.23843622207641602, + 0.08360016345977783, + 0.03924944996833801, + 0.057258885353803635, + 0.08412166684865952, + 0.09090236574411392, + 0.03056921809911728 + ], + "credit_map": [ + 0.08689965307712555, + 0.012242724187672138, + 0.021323187276721, + 0.00881912186741829, + 0.0070310174487531185, + 0.08910171687602997, + 0.015720002353191376, + 0.011824222281575203, + 0.0078037274070084095, + 0.008380940183997154, + 0.006072275340557098, + 0.0031317092943936586, + 0.0043439413420856, + 0.007460275664925575, + 0.003607566934078932, + 0.003604598343372345, + 0.01397554762661457, + 0.007972599007189274, + 0.007063500117510557, + 0.009338312782347202, + 0.008464282378554344, + 0.010344049893319607, + 0.012856199406087399, + 0.008480834774672985, + 0.07818496227264404, + 0.012018437497317791, + 0.09239114075899124, + 0.010645665228366852, + 0.008289643563330173, + 0.012770114466547966, + 0.008724323473870754, + 0.015411936677992344, + 0.010386328212916851, + 0.007387660443782806, + 0.009465553797781467, + 0.007468256633728743, + 0.008370252326130867, + 0.01440209150314331, + 0.01593194715678692, + 0.010188075713813305, + 0.008390389382839203, + 0.0046636792831122875, + 0.007757747545838356, + 0.003424261463806033, + 0.002281960565596819, + 0.0043569342233240604, + 0.00517639284953475, + 0.0031980883795768023, + 0.006520449183881283, + 0.005226224195212126, + 0.00724549638107419, + 0.0060228826478123665, + 0.008339126594364643, + 0.0042458404786884785, + 0.012600102461874485, + 0.007058761548250914, + 0.007171017117798328, + 0.0033306763507425785, + 0.032871074974536896, + 0.0043313391506671906, + 0.0038531392347067595, + 0.017227688804268837, + 0.007574755232781172, + 0.007761974819004536, + 0.00992170162498951, + 0.010766617022454739, + 0.009288613684475422, + 0.012640902772545815, + 0.013049254193902016, + 0.007979685440659523, + 0.015558989718556404, + 0.011696599423885345, + 0.0037332528736442327, + 0.003792579984292388, + 0.0029824175871908665, + 0.0034622035454958677, + 0.004324070177972317, + 0.0023886205162853003, + 0.0060106245800852776, + 0.003875449765473604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.003616936504840851 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368738293647766 + ], + "opposite_logprob_tokens": [ + -3.4086873531341553 + ], + "logprob_reward": -0.03368738293647766, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368738293647766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5241848826408386, + 0.05142569914460182, + 0.03802955895662308, + 0.047929808497428894, + 0.03198206052184105, + 0.0790930986404419, + 0.035955313593149185, + 0.1225251629948616, + 0.016627667471766472, + 0.05224674940109253 + ], + "credit_map": [ + 0.18651659786701202, + 0.02626184932887554, + 0.04238802194595337, + 0.014681676402688026, + 0.013133861124515533, + 0.19045569002628326, + 0.028347399085760117, + 0.022399811074137688, + 0.008413546718657017, + 0.012056193314492702, + 0.006653561722487211, + 0.0041195605881512165, + 0.0045645516365766525, + 0.007592756766825914, + 0.003898937487974763, + 0.004126595798879862, + 0.007428192067891359, + 0.005558171775192022, + 0.00381558109074831, + 0.0038932873867452145, + 0.003981251269578934, + 0.005081627983599901, + 0.004978199489414692, + 0.003293250221759081, + 0.014501508325338364, + 0.003827952081337571, + 0.016169071197509766, + 0.0023215762339532375, + 0.0028783190064132214, + 0.0025857300497591496, + 0.0025981280487030745, + 0.003047526115551591, + 0.004814179148525, + 0.0029415669851005077, + 0.002222549868747592, + 0.003943502437323332, + 0.004045417066663504, + 0.003236391581594944, + 0.006242899689823389, + 0.004535555373877287, + 0.016188016161322594, + 0.009252050891518593, + 0.01386130042374134, + 0.007638942915946245, + 0.006522882729768753, + 0.007675802800804377, + 0.01156656350940466, + 0.006387548986822367, + 0.005709161516278982, + 0.004017879720777273, + 0.006346357986330986, + 0.0031800682190805674, + 0.0032878979109227657, + 0.003929514437913895, + 0.005499064456671476, + 0.0039853742346167564, + 0.015908850356936455, + 0.005817658267915249, + 0.0469912551343441, + 0.007255746517330408, + 0.005153339356184006, + 0.0192109402269125, + 0.010668140836060047, + 0.01151924580335617, + 0.0024172994308173656, + 0.0018682898953557014, + 0.0018252761801704764, + 0.0027275357861071825, + 0.0015328038716688752, + 0.0013447226956486702, + 0.002635742537677288, + 0.002275998005643487, + 0.008084396831691265, + 0.005531401373445988, + 0.009128700010478497, + 0.005867033265531063, + 0.005716125015169382, + 0.004782578907907009, + 0.007841121405363083, + 0.005295394919812679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.03368738293647766 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, tilted close-up focusing on a massive bullfrog?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bullfrog's skin moist, smooth, and speckled with dark green and brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bullfrog crouching on a dark, rain-washed rock with an inflated translucent white vocal sac?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a gurgling stream with distant shore and mountains blurred into inky silhouettes resembling an ink landscape painting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the color scheme primarily black, white, and gray with only dark green and emerald green accents?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870640993118286 + ], + "opposite_logprob_tokens": [ + -1.1370640993118286 + ], + "logprob_reward": -0.3870640993118286, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870640993118286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.507416844367981, + 0.1171887144446373, + 0.015664538368582726, + 0.11347299069166183, + 0.018233509734272957, + 0.04403836652636528, + 0.02226148545742035, + 0.12191471457481384, + 0.011162717826664448, + 0.02864614501595497 + ], + "credit_map": [ + 0.18150703608989716, + 0.02756904810667038, + 0.02586483396589756, + 0.016460606828331947, + 0.016038188710808754, + 0.1907375156879425, + 0.02704930678009987, + 0.02219027653336525, + 0.03612883388996124, + 0.014312044717371464, + 0.008567582815885544, + 0.004101880360394716, + 0.004311602097004652, + 0.04126851633191109, + 0.0032512617763131857, + 0.005246991291642189, + 0.0021167625673115253, + 0.0026028724387288094, + 0.002562975976616144, + 0.0014119006227701902, + 0.002135587390512228, + 0.0010125072440132499, + 0.002048246329650283, + 0.0017736847512423992, + 0.035936448723077774, + 0.006295055150985718, + 0.04343007504940033, + 0.005068211816251278, + 0.005210941657423973, + 0.005816498305648565, + 0.004613397642970085, + 0.007102372124791145, + 0.0019357234705239534, + 0.0017106415471062064, + 0.003920933231711388, + 0.0015295129269361496, + 0.00129464000929147, + 0.0016968821873888373, + 0.003776022233068943, + 0.0023691535461694, + 0.008665108121931553, + 0.0047300937585532665, + 0.008992332965135574, + 0.004344213288277388, + 0.003348954953253269, + 0.004630371928215027, + 0.006021625827997923, + 0.0033056633546948433, + 0.0018627304816618562, + 0.002387865213677287, + 0.005327566992491484, + 0.0025176748167723417, + 0.0026683732867240906, + 0.0019604703411459923, + 0.0031553267035633326, + 0.002381478901952505, + 0.015794502571225166, + 0.0036918926052749157, + 0.038616079837083817, + 0.004895505029708147, + 0.0043174573220312595, + 0.038948994129896164, + 0.005411539692431688, + 0.010238741524517536, + 0.0010827318765223026, + 0.0008774350280873477, + 0.001469933893531561, + 0.0016900121700018644, + 0.0013312442461028695, + 0.0009246869012713432, + 0.0020762584172189236, + 0.0017104147700592875, + 0.003842082107439637, + 0.002839674474671483, + 0.004800016526132822, + 0.0032688670326024294, + 0.003052762011066079, + 0.002382945269346237, + 0.0052922856993973255, + 0.0031675139907747507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.3870640993118286 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_11/index_tensor([8824, 8824])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_11/index_tensor([8824, 8824])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a7cf868b539e12769efe4e616fd6c720e359b12b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_11/index_tensor([8824, 8824])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a chaotic handheld first-person perspective on the deck of a ship, rendered in the style of classic 1980s anime, a frantic artillery bombardment unfolds across a turbulent ocean. The camera, shaky and unstable, captures the scene with raw immediacy. In the distance, a massive wooden galleon, bold black outlines against its dark, weathered timbers, unleashes a fierce attack. The muzzles of the enemy ship's cannons flash with brilliant, hand-painted orange and yellow flames, immediately followed by billowing clouds of thick, opaque white and gray smoke. A black cannonball streaks swiftly across the gloomy sky before crashing into the dark, churning waters near the viewer's ship, erupting in a massive eruption of sea spray and Celtic-hued waters. The ensuing shockwave strikes with brutal force, causing the entire frame to tremble violently as a massive white wave crashes against the hull, sending water cascading down through the damp planks of the deck. The air is thick with a haze of gunpowder and sea fog, and the horizon is dotted with multiple explosions, creating a terrifying and exciting spectacle of high-seas warfare, all defined by the characteristic saturated color palette and dramatic, high-contrast lighting of vintage Japanese anime.", + "seed": 1344, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_11/index_8824_seed_1344_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0816616359166801, + "vlm_logprob": -0.0816616359166801, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.0816616359166801, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2103801965713501 + ], + "opposite_logprob_tokens": [ + -1.71038019657135 + ], + "logprob_reward": -0.2103801965713501, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2103801965713501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43416762351989746, + 0.1464281529188156, + 0.03221336007118225, + 0.04633266106247902, + 0.06170074641704559, + 0.06274455785751343, + 0.024236615747213364, + 0.09732703119516373, + 0.023233413696289062, + 0.0716157928109169 + ], + "credit_map": [ + 0.1290871500968933, + 0.14206911623477936, + 0.013939991593360901, + 0.010404855012893677, + 0.03312765061855316, + 0.017600568011403084, + 0.03793150559067726, + 0.05000678822398186, + 0.010615609586238861, + 0.006232640240341425, + 0.005495193414390087, + 0.012515447102487087, + 0.008643592707812786, + 0.010629705153405666, + 0.04708812013268471, + 0.045207858085632324, + 0.004011067096143961, + 0.003299493808299303, + 0.002486254321411252, + 0.0049280887469649315, + 0.003327024634927511, + 0.002089030109345913, + 0.005662969313561916, + 0.006409435532987118, + 0.005903837736696005, + 0.0029936295468360186, + 0.0031425876077264547, + 0.005414986051619053, + 0.0015081210294738412, + 0.002526759635657072, + 0.01356215588748455, + 0.011280587874352932, + 0.003930020611733198, + 0.00215892237611115, + 0.002542725531384349, + 0.0053856950253248215, + 0.0014990752097219229, + 0.0018080426380038261, + 0.021414864808321, + 0.022961411625146866, + 0.009434057399630547, + 0.006369347218424082, + 0.006776330526918173, + 0.00942693930119276, + 0.003734732046723366, + 0.003996437881141901, + 0.010054406709969044, + 0.012952310033142567, + 0.0018883956363424659, + 0.0017610220238566399, + 0.0017479797825217247, + 0.003986048046499491, + 0.0019745652098208666, + 0.005545382387936115, + 0.003120420267805457, + 0.004212804604321718, + 0.00952232163399458, + 0.006967317778617144, + 0.020716845989227295, + 0.015277907252311707, + 0.005946056451648474, + 0.019196635112166405, + 0.009511086158454418, + 0.010188869200646877, + 0.00245364080183208, + 0.002512057079002261, + 0.0028169387951493263, + 0.003150160424411297, + 0.0027152057737112045, + 0.002642987063154578, + 0.0030046149622648954, + 0.003937809262424707, + 0.005652966443449259, + 0.006094754207879305, + 0.010202333331108093, + 0.009058048017323017, + 0.010778500698506832, + 0.006613303441554308, + 0.014464300125837326, + 0.008751584216952324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.2103801965713501 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994780838489532 + ], + "opposite_logprob_tokens": [ + -1.804947853088379 + ], + "logprob_reward": -0.17994780838489532, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994780838489532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13693000376224518, + 0.08063486218452454, + 0.06918086856603622, + 0.20371338725090027, + 0.06699248403310776, + 0.08237775415182114, + 0.09045043587684631, + 0.12088949233293533, + 0.012362231500446796, + 0.13646849989891052 + ], + "credit_map": [ + 0.02072293497622013, + 0.005927817430347204, + 0.005600482225418091, + 0.006053175311535597, + 0.004273754078894854, + 0.008520616218447685, + 0.035431645810604095, + 0.05039956793189049, + 0.0055915359407663345, + 0.005349334329366684, + 0.004704740364104509, + 0.009246649220585823, + 0.007535713259130716, + 0.008611255325376987, + 0.019491448998451233, + 0.020104186609387398, + 0.0075794910080730915, + 0.0064313230104744434, + 0.005744601599872112, + 0.01114492304623127, + 0.004943094216287136, + 0.004344417713582516, + 0.01608094573020935, + 0.012912067584693432, + 0.011689618229866028, + 0.004711234476417303, + 0.010176093317568302, + 0.025077616795897484, + 0.0056480891071259975, + 0.006302779074758291, + 0.07956115901470184, + 0.060546789318323135, + 0.006036740727722645, + 0.0028329212218523026, + 0.004254226107150316, + 0.00820054393261671, + 0.0027884331066161394, + 0.0034083696082234383, + 0.020398931577801704, + 0.019072309136390686, + 0.009842602536082268, + 0.00766776641830802, + 0.00896022655069828, + 0.012125696055591106, + 0.0061765434220433235, + 0.005385176744312048, + 0.015419084578752518, + 0.016800643876194954, + 0.00783398188650608, + 0.0052221086807549, + 0.004473572131246328, + 0.02066926844418049, + 0.01008391659706831, + 0.013571111485362053, + 0.018811006098985672, + 0.009785464964807034, + 0.009387348778545856, + 0.009591927751898766, + 0.03387182950973511, + 0.014589151367545128, + 0.008147276937961578, + 0.02526104263961315, + 0.008000784553587437, + 0.012040126137435436, + 0.0012838697293773293, + 0.0007335814880207181, + 0.002567081479355693, + 0.0020178183913230896, + 0.0009417459950782359, + 0.0018715832848101854, + 0.0015388928586617112, + 0.0014076590305194259, + 0.009473742917180061, + 0.010545436292886734, + 0.020079998299479485, + 0.018677758052945137, + 0.022432943806052208, + 0.01114691887050867, + 0.027579085901379585, + 0.01653260551393032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.17994780838489532 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076294224709272385 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.0076294224709272385, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076294224709272385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.438698410987854, + 0.11650791764259338, + 0.04069135710597038, + 0.028883375227451324, + 0.018488813191652298, + 0.01888258568942547, + 0.028323577716946602, + 0.1290653496980667, + 0.016037583351135254, + 0.16442112624645233 + ], + "credit_map": [ + 0.10691092908382416, + 0.08650681376457214, + 0.013223151676356792, + 0.00892636924982071, + 0.020429668948054314, + 0.014645504765212536, + 0.07930425554513931, + 0.10875163227319717, + 0.011020520702004433, + 0.009322796016931534, + 0.0076528131030499935, + 0.014164022170007229, + 0.008571639657020569, + 0.024792814627289772, + 0.01913725957274437, + 0.021846037358045578, + 0.004273128695785999, + 0.0038285667542368174, + 0.002914267359301448, + 0.008296141400933266, + 0.0034029693342745304, + 0.002961930586025119, + 0.006901812274008989, + 0.008112535811960697, + 0.0028041526675224304, + 0.0023326107766479254, + 0.0028929447289556265, + 0.0061719175428152084, + 0.0013745547039434314, + 0.0021762766409665346, + 0.0058166091330349445, + 0.005314305424690247, + 0.0029881286900490522, + 0.0021248026750981808, + 0.0018840415868908167, + 0.0023791962303221226, + 0.0007810789393261075, + 0.0008633889374323189, + 0.003900564042851329, + 0.003567609703168273, + 0.0019307925831526518, + 0.0019475860754027963, + 0.0012735468335449696, + 0.003274089191108942, + 0.0010407199151813984, + 0.0008433953626081347, + 0.004228862002491951, + 0.004343592096120119, + 0.0031204218976199627, + 0.002136105438694358, + 0.0020386590622365475, + 0.004962448962032795, + 0.00406791502609849, + 0.0035671310033649206, + 0.003564084880053997, + 0.004866807255893946, + 0.008215508423745632, + 0.004748116713017225, + 0.045460138469934464, + 0.01710858754813671, + 0.00610943092033267, + 0.022277526557445526, + 0.00839188601821661, + 0.016754131764173508, + 0.001246777013875544, + 0.0016457720194011927, + 0.002572556957602501, + 0.0017296976875513792, + 0.001806961023248732, + 0.002822593552991748, + 0.00240395893342793, + 0.0018092641839757562, + 0.012563732452690601, + 0.012373301200568676, + 0.02289295382797718, + 0.02288176119327545, + 0.027167046442627907, + 0.013312441296875477, + 0.03377038985490799, + 0.019459491595625877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0076294224709272385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735478527843952 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735478527843952, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735478527843952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17101189494132996, + 0.15095047652721405, + 0.09270190447568893, + 0.06876253336668015, + 0.07490497827529907, + 0.06925880163908005, + 0.05676666274666786, + 0.04474393650889397, + 0.07864196598529816, + 0.192256897687912 + ], + "credit_map": [ + 0.024528760462999344, + 0.007271682843565941, + 0.008415110409259796, + 0.008201557211577892, + 0.005165742244571447, + 0.009924027137458324, + 0.04463517665863037, + 0.0628698393702507, + 0.009477702900767326, + 0.011756205931305885, + 0.006612374447286129, + 0.011549108661711216, + 0.008208832703530788, + 0.014494191855192184, + 0.04284710809588432, + 0.04600495472550392, + 0.009712493978440762, + 0.009421657770872116, + 0.006650351919233799, + 0.012815228663384914, + 0.011303450912237167, + 0.005554953124374151, + 0.020856540650129318, + 0.016387229785323143, + 0.005609973333775997, + 0.003762284992262721, + 0.0055059161968529224, + 0.010857840068638325, + 0.0033350472804158926, + 0.0038389884866774082, + 0.02044682949781418, + 0.015405653044581413, + 0.012225572019815445, + 0.00957542471587658, + 0.005160022992640734, + 0.009008827619254589, + 0.006062354426831007, + 0.004600176587700844, + 0.011222973465919495, + 0.017049625515937805, + 0.006580245681107044, + 0.009243329055607319, + 0.004946866538375616, + 0.008969048969447613, + 0.005656569730490446, + 0.004971696529537439, + 0.011587567627429962, + 0.017303472384810448, + 0.0054792799055576324, + 0.003365618409588933, + 0.004803816322237253, + 0.006787465885281563, + 0.004442171659320593, + 0.012201759032905102, + 0.011122863739728928, + 0.008563683368265629, + 0.005117605905979872, + 0.0037141924258321524, + 0.009314418770372868, + 0.004464811645448208, + 0.003750384319573641, + 0.009517787955701351, + 0.004481018986552954, + 0.004383716732263565, + 0.006103416904807091, + 0.004426559899002314, + 0.015441399067640305, + 0.0094840032979846, + 0.008996532298624516, + 0.010955924168229103, + 0.011015964671969414, + 0.01221816148608923, + 0.014980221167206764, + 0.01403159648180008, + 0.027186447754502296, + 0.02568093314766884, + 0.03041626513004303, + 0.014741444028913975, + 0.040665313601493835, + 0.024554675444960594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.006735478527843952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615273628383875 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615273628383875, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615273628383875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5124990940093994, + 0.05948544666171074, + 0.01768084056675434, + 0.03239866718649864, + 0.06789085268974304, + 0.018061697483062744, + 0.01663261651992798, + 0.02985662780702114, + 0.014573518186807632, + 0.2309206873178482 + ], + "credit_map": [ + 0.1330372542142868, + 0.11915994435548782, + 0.015526303090155125, + 0.010298041626811028, + 0.0277634896337986, + 0.017726410180330276, + 0.07960949838161469, + 0.10937809944152832, + 0.005618719849735498, + 0.005405502859503031, + 0.0030467144679278135, + 0.00591087294742465, + 0.004522931296378374, + 0.007455151993781328, + 0.013000421226024628, + 0.01452513039112091, + 0.0022804592736065388, + 0.0014375840546563268, + 0.0013239998370409012, + 0.0029940546955913305, + 0.0010918446350842714, + 0.0011370362481102347, + 0.0038341046310961246, + 0.0035817555617541075, + 0.004221349488943815, + 0.0025988491252064705, + 0.002727908082306385, + 0.005814055912196636, + 0.001532903523184359, + 0.0019123703241348267, + 0.006878755521029234, + 0.006712471600621939, + 0.010515185073018074, + 0.007142507471144199, + 0.0035339498426765203, + 0.006565645337104797, + 0.003247406566515565, + 0.0026340526528656483, + 0.01806924119591713, + 0.016182860359549522, + 0.002592835808172822, + 0.0022166904527693987, + 0.0019488453399389982, + 0.002412650501355529, + 0.0012880777940154076, + 0.0012478717835620046, + 0.003275003982707858, + 0.003079721936956048, + 0.0013026887318119407, + 0.0012751190224662423, + 0.0016588879516348243, + 0.002476851223036647, + 0.0015931217931210995, + 0.0037486322689801455, + 0.002669363049790263, + 0.0019079515477642417, + 0.002382344100624323, + 0.001375440857373178, + 0.010119300335645676, + 0.0035740381572395563, + 0.0013561469968408346, + 0.003963063471019268, + 0.0020680720917880535, + 0.005018219817429781, + 0.0012638260377570987, + 0.0014964169822633266, + 0.0022943534422665834, + 0.0016386095667257905, + 0.001423728303052485, + 0.0018714790930971503, + 0.002213660627603531, + 0.002371443435549736, + 0.016466114670038223, + 0.017732491716742516, + 0.03393689543008804, + 0.030478673055768013, + 0.03690974414348602, + 0.019622070714831352, + 0.0477328822016716, + 0.028041796758770943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.003615273628383875 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a chaotic handheld first-person perspective on the deck of a ship, rendered in the style of classic 1980s anime, a frantic artillery bombardment unfolds across a turbulent ocean. The camera, shaky and unstable, captures the scene with raw immediacy. In the distance, a massive wooden galleon, bold black outlines against its dark, weathered timbers, unleashes a fierce attack. The muzzles of the enemy ship's cannons flash with brilliant, hand-painted orange and yellow flames, immediately followed by billowing clouds of thick, opaque white and gray smoke. A black cannonball streaks swiftly across the gloomy sky before crashing into the dark, churning waters near the viewer's ship, erupting in a massive eruption of sea spray and Celtic-hued waters. The ensuing shockwave strikes with brutal force, causing the entire frame to tremble violently as a massive white wave crashes against the hull, sending water cascading down through the damp planks of the deck. The air is thick with a haze of gunpowder and sea fog, and the horizon is dotted with multiple explosions, creating a terrifying and exciting spectacle of high-seas warfare, all defined by the characteristic saturated color palette and dramatic, high-contrast lighting of vintage Japanese anime.", + "seed": 1345, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_11/index_8824_seed_1345_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15130205121822654, + "vlm_logprob": -0.15130205121822654, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.15130205121822654, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049564339220523834 + ], + "opposite_logprob_tokens": [ + -3.4245643615722656 + ], + "logprob_reward": -0.049564339220523834, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.049564339220523834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3719712197780609, + 0.07722719013690948, + 0.11371567845344543, + 0.03682190552353859, + 0.020775238052010536, + 0.03329550102353096, + 0.25687581300735474, + 0.028078004717826843, + 0.027099205181002617, + 0.03414026275277138 + ], + "credit_map": [ + 0.018529823049902916, + 0.008854168467223644, + 0.025610411539673805, + 0.08826224505901337, + 0.007678232155740261, + 0.006854682229459286, + 0.15135270357131958, + 0.06482895463705063, + 0.003379350295290351, + 0.004290127195417881, + 0.009210079908370972, + 0.01855286955833435, + 0.003249367466196418, + 0.0035101270768791437, + 0.022801455110311508, + 0.012233811430633068, + 0.0056605264544487, + 0.002439129399135709, + 0.007235005032271147, + 0.02118104137480259, + 0.002339598722755909, + 0.002288606483489275, + 0.04812174290418625, + 0.02445002645254135, + 0.0017719842726364732, + 0.0017214312683790922, + 0.00554770277813077, + 0.009201863780617714, + 0.0020030969753861427, + 0.0014417109778150916, + 0.009397122077643871, + 0.00573699502274394, + 0.001563464873470366, + 0.0011267946101725101, + 0.0035082544200122356, + 0.00255506276153028, + 0.0016358362045139074, + 0.0008582660229876637, + 0.005941682029515505, + 0.0035858778282999992, + 0.003959354478865862, + 0.0014380577486008406, + 0.004493492655456066, + 0.005835945717990398, + 0.002569429576396942, + 0.0016181799583137035, + 0.010146263055503368, + 0.0032347782980650663, + 0.01681358739733696, + 0.009003669954836369, + 0.08071736246347427, + 0.11098404973745346, + 0.004990710411220789, + 0.008340725675225258, + 0.018021097406744957, + 0.008004603907465935, + 0.004418719559907913, + 0.002660981146618724, + 0.003987543750554323, + 0.003580415388569236, + 0.0033841452095657587, + 0.00350562552921474, + 0.003883089404553175, + 0.0026574856601655483, + 0.0034925888758152723, + 0.003137919818982482, + 0.004691932816058397, + 0.006022343877702951, + 0.0019504616502672434, + 0.0029993748757988214, + 0.0025723972357809544, + 0.0022321853321045637, + 0.0037868130020797253, + 0.006195033900439739, + 0.0027340503875166178, + 0.0036658192984759808, + 0.00217465334571898, + 0.0036510175559669733, + 0.009423194453120232, + 0.0025096822064369917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.049564339220523834 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934017539024353 + ], + "opposite_logprob_tokens": [ + -0.6934017539024353 + ], + "logprob_reward": -0.6934017539024353, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934017539024353, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.5412939190864563, + 0.03268534690141678, + 0.021944064646959305, + 0.017465481534600258, + 0.024672381579875946, + 0.024590039625763893, + 0.2896943688392639, + 0.019575167447328568, + 0.01332999486476183, + 0.014749143272638321 + ], + "credit_map": [ + 0.026185590773820877, + 0.009438058361411095, + 0.03432002663612366, + 0.12037625908851624, + 0.009987732395529747, + 0.007338497322052717, + 0.2352273166179657, + 0.09842050075531006, + 0.0017381733050569892, + 0.002020237734541297, + 0.0044001080095767975, + 0.007183725945651531, + 0.0011357753537595272, + 0.0012786021688953042, + 0.009151831269264221, + 0.005776895675808191, + 0.0016603284748271108, + 0.001010351232253015, + 0.0018082520691677928, + 0.004259822890162468, + 0.0009416927350685, + 0.001395813887938857, + 0.007083274889737368, + 0.0037845305632799864, + 0.0011741053313016891, + 0.001031035091727972, + 0.002125416649505496, + 0.003821594174951315, + 0.0013301161816343665, + 0.001203219871968031, + 0.003678110893815756, + 0.0031018846202641726, + 0.0022136070765554905, + 0.0015021192375570536, + 0.005149537231773138, + 0.00404870742931962, + 0.0011883687693625689, + 0.0015320135280489922, + 0.0055275182239711285, + 0.0035105119459331036, + 0.00268965819850564, + 0.0008477414958178997, + 0.0042584738694131374, + 0.00468203891068697, + 0.0009636399336159229, + 0.000783184077590704, + 0.007250348571687937, + 0.003114955732598901, + 0.019651278853416443, + 0.00855791475623846, + 0.0852106362581253, + 0.12987226247787476, + 0.005145902745425701, + 0.006975686177611351, + 0.024424875155091286, + 0.009855857118964195, + 0.0029403481166809797, + 0.0026608146727085114, + 0.002721987897530198, + 0.003011199878528714, + 0.002086296444758773, + 0.002736001741141081, + 0.002032103016972542, + 0.001386418822221458, + 0.0014951219782233238, + 0.0015529304509982467, + 0.002640094608068466, + 0.002676284871995449, + 0.0012555897701531649, + 0.0014248151564970613, + 0.0010206387378275394, + 0.0012645203387364745, + 0.001530611189082265, + 0.0019289744086563587, + 0.0016288980841636658, + 0.002245853189378977, + 0.0022392370738089085, + 0.0018269866704940796, + 0.0021219230256974697, + 0.0012266612611711025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.6934017539024353 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003197915619239211 + ], + "opposite_logprob_tokens": [ + -5.753198146820068 + ], + "logprob_reward": -0.003197915619239211, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003197915619239211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1274224817752838, + 0.10910311341285706, + 0.07033821940422058, + 0.04693980515003204, + 0.07120149582624435, + 0.03349502757191658, + 0.47358238697052, + 0.009701676666736603, + 0.01910282112658024, + 0.03911295905709267 + ], + "credit_map": [ + 0.006581773515790701, + 0.006916197948157787, + 0.009126856923103333, + 0.030272571370005608, + 0.005897285416722298, + 0.004435379058122635, + 0.04381347820162773, + 0.02037893794476986, + 0.005976686254143715, + 0.005966749042272568, + 0.01155560091137886, + 0.024571647867560387, + 0.005914825014770031, + 0.0054654451087117195, + 0.031333524733781815, + 0.01831863820552826, + 0.004552056547254324, + 0.004078949335962534, + 0.008726531639695168, + 0.01392280776053667, + 0.003845600178465247, + 0.0038666841574013233, + 0.016669468954205513, + 0.014676123857498169, + 0.003247227054089308, + 0.0019731048960238695, + 0.0052378615364432335, + 0.011089377105236053, + 0.0016709552146494389, + 0.002896784571930766, + 0.01370148453861475, + 0.007123010698705912, + 0.005729014985263348, + 0.004264469724148512, + 0.010651711374521255, + 0.01035646628588438, + 0.003896040841937065, + 0.004280549008399248, + 0.02099830098450184, + 0.011024938896298409, + 0.0037375101819634438, + 0.0009866462787613273, + 0.0031619169749319553, + 0.008077523671090603, + 0.0015716045163571835, + 0.0011358901392668486, + 0.012369615025818348, + 0.002454322064295411, + 0.030741263180971146, + 0.015025516971945763, + 0.14873023331165314, + 0.20796383917331696, + 0.007789207622408867, + 0.013301282189786434, + 0.036203596740961075, + 0.013827454298734665, + 0.0011461831163614988, + 0.0020098804961889982, + 0.001279981923289597, + 0.0014015374472364783, + 0.0009865107713267207, + 0.0012267756974324584, + 0.0008149241912178695, + 0.0008358829654753208, + 0.0016890596598386765, + 0.0026322125922888517, + 0.0025242012925446033, + 0.0028136747423559427, + 0.002180389827117324, + 0.002362708793953061, + 0.0019477076129987836, + 0.002952866954728961, + 0.0034414317924529314, + 0.004461762495338917, + 0.0032551295589655638, + 0.004797282163053751, + 0.005300919059664011, + 0.003801387967541814, + 0.010820967145264149, + 0.0032340814359486103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.003197915619239211 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067327553406357765 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.0067327553406357765, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067327553406357765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046539757400751114, + 0.015719449147582054, + 0.033836256712675095, + 0.018529243767261505, + 0.016100319102406502, + 0.045473504811525345, + 0.7936258316040039, + 0.015019817277789116, + 0.005656025838106871, + 0.009499789215624332 + ], + "credit_map": [ + 0.001964820781722665, + 0.002282456960529089, + 0.004478336777538061, + 0.011422129347920418, + 0.002124352613463998, + 0.0016821742756292224, + 0.014567921869456768, + 0.008017566055059433, + 0.0012736818753182888, + 0.0012015613028779626, + 0.0017336781602352858, + 0.0035215069074183702, + 0.0012704627588391304, + 0.0009389574988745153, + 0.0027360734529793262, + 0.0030435281805694103, + 0.0014253403060138226, + 0.0012253363383933902, + 0.0019054211443290114, + 0.005997775122523308, + 0.0010915356688201427, + 0.0011776420287787914, + 0.013698982074856758, + 0.007314222864806652, + 0.0008798532653599977, + 0.000924331194255501, + 0.0024027936160564423, + 0.005502972286194563, + 0.0008866388234309852, + 0.0009242871310561895, + 0.004247421398758888, + 0.0027609437238425016, + 0.001140190870501101, + 0.0007257055258378386, + 0.002491843421012163, + 0.0023249504156410694, + 0.0009408861515112221, + 0.0007520568324252963, + 0.005433764774352312, + 0.0022909224499017, + 0.005286616273224354, + 0.001274934853427112, + 0.0046972911804914474, + 0.010574901476502419, + 0.0017738626338541508, + 0.0014257520670071244, + 0.01709088496863842, + 0.0033492641523480415, + 0.05272504314780235, + 0.02284836396574974, + 0.23914793133735657, + 0.356317400932312, + 0.012906117364764214, + 0.01801474392414093, + 0.0661369115114212, + 0.025529367849230766, + 0.0018859899137169123, + 0.002386817242950201, + 0.0022524246014654636, + 0.0021232913713902235, + 0.001883100951090455, + 0.0015224067028611898, + 0.001674586907029152, + 0.0012912009842693806, + 0.000710563559550792, + 0.00069709203671664, + 0.000787534168921411, + 0.0009993441635742784, + 0.0006128710228949785, + 0.0007621520198881626, + 0.0006048191571608186, + 0.00048164970939978957, + 0.0007978538051247597, + 0.0009809329640120268, + 0.000827974290587008, + 0.0013716930989176035, + 0.0008684871136210859, + 0.0009004842140711844, + 0.002909647999331355, + 0.0008427168941125274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0067327553406357765 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036134920082986355 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.0036134920082986355, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036134920082986355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15094101428985596, + 0.0372677817940712, + 0.028890926390886307, + 0.016391126438975334, + 0.0214559156447649, + 0.014328151009976864, + 0.6850851774215698, + 0.008915998972952366, + 0.009194118902087212, + 0.027529843151569366 + ], + "credit_map": [ + 0.007957084104418755, + 0.005214091390371323, + 0.00998777151107788, + 0.03540569171309471, + 0.003788947593420744, + 0.0047020819038152695, + 0.05828080326318741, + 0.025604519993066788, + 0.0028114926535636187, + 0.0028347275219857693, + 0.0037941704504191875, + 0.007280309218913317, + 0.0041654473170638084, + 0.0023201913572847843, + 0.007722913753241301, + 0.006338526960462332, + 0.0016617161454632878, + 0.0011143471347168088, + 0.002272672951221466, + 0.005755146034061909, + 0.0011651552049443126, + 0.0012197063770145178, + 0.010391025803983212, + 0.005311151966452599, + 0.0012652118457481265, + 0.0008619753643870354, + 0.0021072416566312313, + 0.0034503722563385963, + 0.0011968265753239393, + 0.0013327032793313265, + 0.003176337806507945, + 0.0030004559084773064, + 0.0013509036507457495, + 0.0012744024861603975, + 0.0035153310745954514, + 0.0034399295691400766, + 0.0013290704227983952, + 0.0015180421760305762, + 0.00607546791434288, + 0.002952766604721546, + 0.0015979352174326777, + 0.0004991278983652592, + 0.002481977455317974, + 0.003019415307790041, + 0.0006368590984493494, + 0.0005859341472387314, + 0.003604929195716977, + 0.0019019708270207047, + 0.04474330320954323, + 0.019243335351347923, + 0.20657984912395477, + 0.3070651590824127, + 0.011081255972385406, + 0.016001395881175995, + 0.05902775377035141, + 0.021343106403946877, + 0.001190546085126698, + 0.0012083589099347591, + 0.001053722808137536, + 0.001292460598051548, + 0.0011817397316917777, + 0.0009453201782889664, + 0.0011222446337342262, + 0.0009216054459102452, + 0.0011434925254434347, + 0.0012708852300420403, + 0.0012889682548120618, + 0.0017988854087889194, + 0.0009048472275026143, + 0.0011022891849279404, + 0.0008941446430981159, + 0.0007906061364337802, + 0.0025888998061418533, + 0.004363488871604204, + 0.0026630216743797064, + 0.0037903718184679747, + 0.002971469657495618, + 0.003378759603947401, + 0.005476394202560186, + 0.0022974354214966297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0036134920082986355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a chaotic handheld first-person perspective on the deck of a ship, rendered in the style of classic 1980s anime, a frantic artillery bombardment unfolds across a turbulent ocean. The camera, shaky and unstable, captures the scene with raw immediacy. In the distance, a massive wooden galleon, bold black outlines against its dark, weathered timbers, unleashes a fierce attack. The muzzles of the enemy ship's cannons flash with brilliant, hand-painted orange and yellow flames, immediately followed by billowing clouds of thick, opaque white and gray smoke. A black cannonball streaks swiftly across the gloomy sky before crashing into the dark, churning waters near the viewer's ship, erupting in a massive eruption of sea spray and Celtic-hued waters. The ensuing shockwave strikes with brutal force, causing the entire frame to tremble violently as a massive white wave crashes against the hull, sending water cascading down through the damp planks of the deck. The air is thick with a haze of gunpowder and sea fog, and the horizon is dotted with multiple explosions, creating a terrifying and exciting spectacle of high-seas warfare, all defined by the characteristic saturated color palette and dramatic, high-contrast lighting of vintage Japanese anime.", + "seed": 1346, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_11/index_8824_seed_1346_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17618443807587028, + "vlm_logprob": -0.17618443807587028, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.17618443807587028, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43667471408843994 + ], + "opposite_logprob_tokens": [ + -1.06167471408844 + ], + "logprob_reward": -0.43667471408843994, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.43667471408843994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.348302960395813, + 0.13700903952121735, + 0.11899542063474655, + 0.05963355675339699, + 0.05891552194952965, + 0.06308329850435257, + 0.11853009462356567, + 0.023459888994693756, + 0.039601586759090424, + 0.032468631863594055 + ], + "credit_map": [ + 0.023125074803829193, + 0.044855181127786636, + 0.04127936437726021, + 0.03645890951156616, + 0.03997453674674034, + 0.034032903611660004, + 0.08713699132204056, + 0.04144001007080078, + 0.009497350081801414, + 0.005361464805901051, + 0.041245050728321075, + 0.014395084232091904, + 0.009029507637023926, + 0.011255105957388878, + 0.009159774519503117, + 0.03706568479537964, + 0.012584027834236622, + 0.008171924389898777, + 0.022251000627875328, + 0.02058318629860878, + 0.011752558872103691, + 0.0095541225746274, + 0.014315239153802395, + 0.019783353433012962, + 0.004988484084606171, + 0.004692364484071732, + 0.011309675872325897, + 0.007497549522668123, + 0.009455791674554348, + 0.004399892408400774, + 0.0076339999213814735, + 0.009655790403485298, + 0.0048972973600029945, + 0.008441662415862083, + 0.008700359612703323, + 0.004238997120410204, + 0.006628703325986862, + 0.005981325637549162, + 0.01571865752339363, + 0.004308516625314951, + 0.004561905283480883, + 0.007457824889570475, + 0.004952793940901756, + 0.005119501147419214, + 0.01312794629484415, + 0.006299952976405621, + 0.011609521694481373, + 0.009953850880265236, + 0.01223283726722002, + 0.011678911745548248, + 0.012763611041009426, + 0.01297947857528925, + 0.012729398906230927, + 0.01912795938551426, + 0.016606301069259644, + 0.02041158452630043, + 0.0035215406678617, + 0.003148801391944289, + 0.0028610755689442158, + 0.0025230655446648598, + 0.002619235310703516, + 0.003067736281082034, + 0.0032024490647017956, + 0.002515981672331691, + 0.005338429473340511, + 0.004745790734887123, + 0.004406164400279522, + 0.004959231708198786, + 0.0093791913241148, + 0.005003188271075487, + 0.0033040305133908987, + 0.0024655540473759174, + 0.002553155180066824, + 0.005443686619400978, + 0.004104045685380697, + 0.004109564702957869, + 0.00441187247633934, + 0.005272633861750364, + 0.0037672475446015596, + 0.0028064236976206303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.43667471408843994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289517402648926 + ], + "opposite_logprob_tokens": [ + -1.0539517402648926 + ], + "logprob_reward": -0.4289517402648926, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289517402648926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11182413250207901, + 0.1721700131893158, + 0.07801224291324615, + 0.07140009105205536, + 0.08353231102228165, + 0.06859780848026276, + 0.24956898391246796, + 0.03360883519053459, + 0.06237948685884476, + 0.06890604645013809 + ], + "credit_map": [ + 0.00990586169064045, + 0.012107320129871368, + 0.013807883486151695, + 0.014926171861588955, + 0.016721230000257492, + 0.013011293485760689, + 0.020488876849412918, + 0.010855497792363167, + 0.01917347125709057, + 0.007971866056323051, + 0.04569033160805702, + 0.026181671768426895, + 0.008835436776280403, + 0.010581565089523792, + 0.010923728346824646, + 0.042811937630176544, + 0.00840329471975565, + 0.004851533565670252, + 0.016199767589569092, + 0.013463915325701237, + 0.008140320889651775, + 0.006381637416779995, + 0.007203331682831049, + 0.01336844451725483, + 0.0058167255483567715, + 0.0047610546462237835, + 0.011900469660758972, + 0.009316807612776756, + 0.008076509460806847, + 0.004278264939785004, + 0.01426326110959053, + 0.01298699900507927, + 0.006458713207393885, + 0.006236673332750797, + 0.020241454243659973, + 0.005521717946976423, + 0.007703468669205904, + 0.009869396686553955, + 0.01632704958319664, + 0.011173835024237633, + 0.006522763520479202, + 0.008151497691869736, + 0.006962589919567108, + 0.006534887943416834, + 0.017648352310061455, + 0.008758229203522205, + 0.008530164137482643, + 0.005489326547831297, + 0.02876945026218891, + 0.036349620670080185, + 0.036433424800634384, + 0.027755120769143105, + 0.0246697086840868, + 0.03760377690196037, + 0.028442855924367905, + 0.029545027762651443, + 0.0030189217068254948, + 0.0035260221920907497, + 0.003454851685091853, + 0.006717443931847811, + 0.0037712084595113993, + 0.003887733444571495, + 0.005022605415433645, + 0.004210048355162144, + 0.005528565496206284, + 0.008081888779997826, + 0.00811525247991085, + 0.005872375797480345, + 0.01801588386297226, + 0.00689571350812912, + 0.0053920806385576725, + 0.004477722570300102, + 0.007191617041826248, + 0.007199506275355816, + 0.0073886835016310215, + 0.006139071192592382, + 0.008254232816398144, + 0.02027573250234127, + 0.006190843880176544, + 0.006266360636800528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.4289517402648926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951066501438618 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005951066501438618, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951066501438618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2531583905220032, + 0.1424570083618164, + 0.09473510086536407, + 0.12696895003318787, + 0.0933743491768837, + 0.08832290023565292, + 0.07866685092449188, + 0.037999507039785385, + 0.03379157558083534, + 0.05052543431520462 + ], + "credit_map": [ + 0.012999922968447208, + 0.03888159245252609, + 0.027715889737010002, + 0.02760961279273033, + 0.02021143026649952, + 0.014127657748758793, + 0.0806347206234932, + 0.030977556481957436, + 0.014552338048815727, + 0.007654009386897087, + 0.03756719455122948, + 0.011498013511300087, + 0.01687309704720974, + 0.009159858338534832, + 0.008500971831381321, + 0.03665151819586754, + 0.007568467874079943, + 0.007043963298201561, + 0.01605728454887867, + 0.011869965121150017, + 0.01598983444273472, + 0.0061517320573329926, + 0.013553141616284847, + 0.016500698402523994, + 0.013038280420005322, + 0.017727190628647804, + 0.0225679948925972, + 0.013226134702563286, + 0.012075627222657204, + 0.007902532815933228, + 0.02074216678738594, + 0.019689006730914116, + 0.010446809232234955, + 0.014414461329579353, + 0.015811221674084663, + 0.006716488394886255, + 0.01033097505569458, + 0.010122565552592278, + 0.01740376651287079, + 0.00812805537134409, + 0.0054888068698346615, + 0.0067661358043551445, + 0.004122658632695675, + 0.004515431821346283, + 0.010125033557415009, + 0.005191922187805176, + 0.026948899030685425, + 0.025164008140563965, + 0.008605490438640118, + 0.010366152040660381, + 0.012582377530634403, + 0.007889857515692711, + 0.008085466921329498, + 0.010568968951702118, + 0.010090908966958523, + 0.010477625764906406, + 0.004980735946446657, + 0.0043893977999687195, + 0.003672530874609947, + 0.005030851811170578, + 0.004776993300765753, + 0.005012239329516888, + 0.0056423912756145, + 0.004494363907724619, + 0.003297233022749424, + 0.004886596463620663, + 0.003960175905376673, + 0.0027508095372468233, + 0.010340383276343346, + 0.003525278763845563, + 0.0031428365036845207, + 0.0018882597796618938, + 0.005144523456692696, + 0.005900820717215538, + 0.007266127038747072, + 0.00728611322119832, + 0.0056251417845487595, + 0.00860736146569252, + 0.005890520755201578, + 0.004804820753633976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.005951066501438618 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250474438071251 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.005250474438071251, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250474438071251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21373064815998077, + 0.11023883521556854, + 0.057245299220085144, + 0.07437512278556824, + 0.12783877551555634, + 0.057382334023714066, + 0.057923708111047745, + 0.027673695236444473, + 0.04095021262764931, + 0.23264139890670776 + ], + "credit_map": [ + 0.013762376271188259, + 0.030480535700917244, + 0.024784833192825317, + 0.026581384241580963, + 0.016596345230937004, + 0.017326215282082558, + 0.05939275771379471, + 0.024806195870041847, + 0.005551154259592295, + 0.0038645185995846987, + 0.04004780948162079, + 0.011288189329206944, + 0.004705224186182022, + 0.0038881502114236355, + 0.005126230884343386, + 0.0357675701379776, + 0.004423331934958696, + 0.003682461567223072, + 0.010018527507781982, + 0.008126115426421165, + 0.007837898097932339, + 0.005239345133304596, + 0.008213643915951252, + 0.009703978896141052, + 0.006236719898879528, + 0.006827543023973703, + 0.015323972329497337, + 0.007098272908478975, + 0.007569358684122562, + 0.005048452410846949, + 0.01223001815378666, + 0.014040783978998661, + 0.018445249646902084, + 0.014686035923659801, + 0.018290042877197266, + 0.011655837297439575, + 0.02082211710512638, + 0.0209918525069952, + 0.010091574862599373, + 0.01285606250166893, + 0.0034978887997567654, + 0.00486537953838706, + 0.0029038579668849707, + 0.0031598496716469526, + 0.006768617779016495, + 0.004817764274775982, + 0.016221225261688232, + 0.015147754922509193, + 0.005821870639920235, + 0.005023219622671604, + 0.007039577234536409, + 0.007940808311104774, + 0.003953505773097277, + 0.006836751941591501, + 0.012443526647984982, + 0.00886444840580225, + 0.002993952948600054, + 0.001963413320481777, + 0.0030614063143730164, + 0.004657600075006485, + 0.004850784782320261, + 0.003117674496024847, + 0.0036396111827343702, + 0.0033892502542585135, + 0.005294319707900286, + 0.004651571158319712, + 0.002827898832038045, + 0.004892570897936821, + 0.009113727137446404, + 0.005636213347315788, + 0.0054062288254499435, + 0.003127683885395527, + 0.014300674200057983, + 0.06360960751771927, + 0.01457175426185131, + 0.02569648250937462, + 0.06335528939962387, + 0.013789708726108074, + 0.022793490439653397, + 0.014524383470416069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.005250474438071251 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094195086508989 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094195086508989, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094195086508989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15168523788452148, + 0.08882109820842743, + 0.052326999604701996, + 0.12592095136642456, + 0.1815924346446991, + 0.11035376787185669, + 0.10207749903202057, + 0.05222780629992485, + 0.027587715536355972, + 0.10740652680397034 + ], + "credit_map": [ + 0.007068409584462643, + 0.02048732340335846, + 0.01665823720395565, + 0.020261267200112343, + 0.01439280342310667, + 0.009015016257762909, + 0.04594622179865837, + 0.01785595342516899, + 0.007285710424184799, + 0.004995863884687424, + 0.0236415583640337, + 0.012174995616078377, + 0.0095328064635396, + 0.006528580095618963, + 0.006541115231812, + 0.018120473250746727, + 0.006110436283051968, + 0.004122116137295961, + 0.007481880020350218, + 0.007266253232955933, + 0.007153542712330818, + 0.003064093878492713, + 0.008583622053265572, + 0.008545052260160446, + 0.009912898764014244, + 0.023265238851308823, + 0.018454359844326973, + 0.01253828126937151, + 0.025632495060563087, + 0.008329832926392555, + 0.008965731598436832, + 0.01882210001349449, + 0.023254143074154854, + 0.022299503907561302, + 0.023724265396595, + 0.01737062819302082, + 0.028978344053030014, + 0.027607863768935204, + 0.02341485768556595, + 0.014942816458642483, + 0.007519385777413845, + 0.01013468112796545, + 0.007214040961116552, + 0.008804941549897194, + 0.022481119260191917, + 0.010563424788415432, + 0.02412816882133484, + 0.01950800232589245, + 0.010664327070116997, + 0.009773428551852703, + 0.02148124761879444, + 0.012304063886404037, + 0.008009656332433224, + 0.008826355449855328, + 0.016658959910273552, + 0.01435945462435484, + 0.006645179819315672, + 0.009615783579647541, + 0.0040245214477181435, + 0.005272922571748495, + 0.009139303117990494, + 0.007511727046221495, + 0.006519543007016182, + 0.003498826175928116, + 0.0032703590113669634, + 0.0029131604824215174, + 0.003327213227748871, + 0.003295944072306156, + 0.004634242970496416, + 0.004105663392692804, + 0.0031370746437460184, + 0.002904058201238513, + 0.01047704927623272, + 0.016147080808877945, + 0.009724187664687634, + 0.012419401668012142, + 0.012872892431914806, + 0.024739347398281097, + 0.012403263710439205, + 0.008623306639492512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.004094195086508989 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a chaotic handheld first-person perspective on the deck of a ship, rendered in the style of classic 1980s anime, a frantic artillery bombardment unfolds across a turbulent ocean. The camera, shaky and unstable, captures the scene with raw immediacy. In the distance, a massive wooden galleon, bold black outlines against its dark, weathered timbers, unleashes a fierce attack. The muzzles of the enemy ship's cannons flash with brilliant, hand-painted orange and yellow flames, immediately followed by billowing clouds of thick, opaque white and gray smoke. A black cannonball streaks swiftly across the gloomy sky before crashing into the dark, churning waters near the viewer's ship, erupting in a massive eruption of sea spray and Celtic-hued waters. The ensuing shockwave strikes with brutal force, causing the entire frame to tremble violently as a massive white wave crashes against the hull, sending water cascading down through the damp planks of the deck. The air is thick with a haze of gunpowder and sea fog, and the horizon is dotted with multiple explosions, creating a terrifying and exciting spectacle of high-seas warfare, all defined by the characteristic saturated color palette and dramatic, high-contrast lighting of vintage Japanese anime.", + "seed": 1347, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_11/index_8824_seed_1347_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12646962041035295, + "vlm_logprob": -0.12646962041035295, + "vlm_yes_no_margin": 3.474999988079071, + "vlm_token_credit": -0.12646962041035295, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26462215185165405 + ], + "opposite_logprob_tokens": [ + -1.5146222114562988 + ], + "logprob_reward": -0.26462215185165405, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.26462215185165405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04561701789498329, + 0.036418233066797256, + 0.012763057835400105, + 0.01484345830976963, + 0.04010729864239693, + 0.061389192938804626, + 0.057807762175798416, + 0.008630477823317051, + 0.20268473029136658, + 0.5197387337684631 + ], + "credit_map": [ + 0.007484584581106901, + 0.006815578322857618, + 0.006738161202520132, + 0.00389782409183681, + 0.005254056770354509, + 0.005147916730493307, + 0.004628696944564581, + 0.0056502022780478, + 0.005355053581297398, + 0.007088836282491684, + 0.004007078241556883, + 0.0033085420727729797, + 0.005282434169203043, + 0.00566877843812108, + 0.003310137428343296, + 0.002397376112639904, + 0.0011476216604933143, + 0.0023734658025205135, + 0.00165198661852628, + 0.0013703369768336415, + 0.0021036891266703606, + 0.0014132503420114517, + 0.0015529304509982467, + 0.0011497782543301582, + 0.0012052017264068127, + 0.0010715576354414225, + 0.0024344632402062416, + 0.002402716549113393, + 0.0016245594015344977, + 0.0008888925076462328, + 0.0029490834567695856, + 0.002266985597088933, + 0.0036003596615046263, + 0.0030047863256186247, + 0.0038474907632917166, + 0.005012300796806812, + 0.0055617038160562515, + 0.0038047845009714365, + 0.00640474259853363, + 0.008871135301887989, + 0.010316731408238411, + 0.0052083986811339855, + 0.0032699855510145426, + 0.00473679369315505, + 0.008357537910342216, + 0.008703760802745819, + 0.011405968107283115, + 0.009390024468302727, + 0.006689238827675581, + 0.008094055578112602, + 0.006223269738256931, + 0.008780330419540405, + 0.006016308441758156, + 0.007419591769576073, + 0.007535949349403381, + 0.007049023173749447, + 0.0008755445596762002, + 0.001233640592545271, + 0.001026585465297103, + 0.0017934029456228018, + 0.0009627957479096949, + 0.0009466851479373872, + 0.001119214459322393, + 0.000672609603498131, + 0.03181222453713417, + 0.015789320692420006, + 0.07500621676445007, + 0.019198603928089142, + 0.020860830321907997, + 0.007816490717232227, + 0.015658406540751457, + 0.016542650759220123, + 0.02001654915511608, + 0.008155452087521553, + 0.02102285996079445, + 0.04807169735431671, + 0.01678589917719364, + 0.008579072542488575, + 0.2162235826253891, + 0.18088366091251373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.26462215185165405 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487265706062317 + ], + "opposite_logprob_tokens": [ + -1.223726511001587 + ], + "logprob_reward": -0.3487265706062317, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3487265706062317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028066907078027725, + 0.03365905582904816, + 0.024619942530989647, + 0.04566635563969612, + 0.03404584527015686, + 0.05483403429389, + 0.015932921320199966, + 0.026711495593190193, + 0.1258150190114975, + 0.6106484532356262 + ], + "credit_map": [ + 0.0034852703101933002, + 0.00304117682389915, + 0.0034134725574404, + 0.0028654588386416435, + 0.003448238829150796, + 0.003588919062167406, + 0.004019177984446287, + 0.004205191973596811, + 0.005205008666962385, + 0.0063214609399437904, + 0.004183567129075527, + 0.0024832964409142733, + 0.004877012223005295, + 0.00547203840687871, + 0.003020867006853223, + 0.0020958082750439644, + 0.0019334577955305576, + 0.004859217442572117, + 0.003080737544223666, + 0.0026489165611565113, + 0.003549395129084587, + 0.0030165021307766438, + 0.003751434851437807, + 0.0017802828224375844, + 0.004015124868601561, + 0.005633361171931028, + 0.00744672492146492, + 0.005903718527406454, + 0.007000450044870377, + 0.003733780002221465, + 0.006900110747665167, + 0.005033084191381931, + 0.002730597974732518, + 0.003248609369620681, + 0.00399739108979702, + 0.005413457285612822, + 0.003778162645176053, + 0.004803624469786882, + 0.004621893633157015, + 0.0054521095007658005, + 0.004484752658754587, + 0.002944627311080694, + 0.00347748096100986, + 0.004726038780063391, + 0.00619164016097784, + 0.005685527343302965, + 0.01711413636803627, + 0.010209832340478897, + 0.0017249849624931812, + 0.0018591736443340778, + 0.0019426846411079168, + 0.00252394936978817, + 0.0012516581919044256, + 0.002291255397722125, + 0.002041444880887866, + 0.0022977711632847786, + 0.0033342852257192135, + 0.004851989448070526, + 0.0026646675541996956, + 0.00423227297142148, + 0.0038621535059064627, + 0.0024995869025588036, + 0.0026987639721482992, + 0.002567775547504425, + 0.02031833305954933, + 0.00871968548744917, + 0.04751545563340187, + 0.009455912746489048, + 0.012288854457437992, + 0.004833936225622892, + 0.011430705897510052, + 0.011252129450440407, + 0.025457778945565224, + 0.01176435500383377, + 0.027519971132278442, + 0.059400979429483414, + 0.020148059353232384, + 0.009083644486963749, + 0.252384752035141, + 0.20488886535167694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.3487265706062317 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739267613738775 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006739267613738775, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739267613738775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21137785911560059, + 0.0147478012368083, + 0.07191655784845352, + 0.014429253526031971, + 0.055987682193517685, + 0.10222136229276657, + 0.08657485246658325, + 0.03111911192536354, + 0.15795761346817017, + 0.2536678612232208 + ], + "credit_map": [ + 0.034942034631967545, + 0.033860962837934494, + 0.02000316232442856, + 0.02766251564025879, + 0.02188875526189804, + 0.02080542780458927, + 0.030411213636398315, + 0.021803779527544975, + 0.00184622744563967, + 0.0020558335818350315, + 0.001880116295069456, + 0.0022217249497771263, + 0.0016297429101541638, + 0.001665799762122333, + 0.0019463325152173638, + 0.0015020242426544428, + 0.006582998670637608, + 0.016002275049686432, + 0.00886000506579876, + 0.008137285709381104, + 0.009073112159967422, + 0.008690377697348595, + 0.009692358784377575, + 0.004878145642578602, + 0.0012529940577223897, + 0.0016167290741577744, + 0.0026408517733216286, + 0.0018189449328929186, + 0.0017077198717743158, + 0.0011646229540929198, + 0.0019834875129163265, + 0.0022439032327383757, + 0.006132418289780617, + 0.0026986636221408844, + 0.0022598630748689175, + 0.0026642344892024994, + 0.013237531296908855, + 0.003400399349629879, + 0.00912492722272873, + 0.016469646245241165, + 0.009233246557414532, + 0.0049627008847892284, + 0.0072404444217681885, + 0.0153411366045475, + 0.008513581939041615, + 0.014102444052696228, + 0.030307605862617493, + 0.012520200572907925, + 0.009680130518972874, + 0.013147602789103985, + 0.009114088490605354, + 0.011280759237706661, + 0.011010050773620605, + 0.011256339959800243, + 0.011139191687107086, + 0.009946689009666443, + 0.003319383831694722, + 0.0033623813651502132, + 0.003832015208899975, + 0.004135416354984045, + 0.004220744129270315, + 0.0034916342701762915, + 0.004381600301712751, + 0.004375939257442951, + 0.014619190245866776, + 0.017040647566318512, + 0.0170210599899292, + 0.01716184988617897, + 0.036343324929475784, + 0.010645734146237373, + 0.026474634185433388, + 0.018651161342859268, + 0.008801667019724846, + 0.007229319307953119, + 0.010201008059084415, + 0.02192535623908043, + 0.011366592720150948, + 0.006085987202823162, + 0.10036657005548477, + 0.08769134432077408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.006739267613738775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624808698892593 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624808698892593, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624808698892593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04200545325875282, + 0.0107969269156456, + 0.025171201676130295, + 0.05333653464913368, + 0.28794804215431213, + 0.09613823145627975, + 0.02597789652645588, + 0.029860690236091614, + 0.13917461037635803, + 0.28959041833877563 + ], + "credit_map": [ + 0.00730745168402791, + 0.007632576394826174, + 0.0049620578065514565, + 0.004408367443829775, + 0.004589940421283245, + 0.004947919864207506, + 0.004086984787136316, + 0.004070156719535589, + 0.0013976836344227195, + 0.0017067723674699664, + 0.0013594810152426362, + 0.0013880057958886027, + 0.0015288956928998232, + 0.0011454977793619037, + 0.0011590429348871112, + 0.0011115485103800893, + 0.0031144381500780582, + 0.004300793167203665, + 0.003313584951683879, + 0.0028250510804355145, + 0.003960565663874149, + 0.0029811770655214787, + 0.002718838630244136, + 0.0019567522685974836, + 0.004636980593204498, + 0.0040449174121022224, + 0.00958296563476324, + 0.010556584224104881, + 0.0072110542096197605, + 0.0027287073899060488, + 0.008251344785094261, + 0.006323978770524263, + 0.03163398429751396, + 0.01111902017146349, + 0.0073519619181752205, + 0.01060130912810564, + 0.07201468199491501, + 0.016539977863430977, + 0.04873618483543396, + 0.08995091915130615, + 0.0251530259847641, + 0.0060652983374893665, + 0.0053800721652805805, + 0.009427239187061787, + 0.013722850941121578, + 0.009452403523027897, + 0.01677689142525196, + 0.010160457342863083, + 0.0025856683496385813, + 0.0033692552242428064, + 0.0022222530096769333, + 0.004750915803015232, + 0.003403350478038192, + 0.003269931534305215, + 0.003373746294528246, + 0.0030027776956558228, + 0.0026755149010568857, + 0.002821102272719145, + 0.00408089067786932, + 0.005492738448083401, + 0.003816033946350217, + 0.0036590059753507376, + 0.004221752285957336, + 0.0030936524271965027, + 0.022063041105866432, + 0.009422537870705128, + 0.05483913794159889, + 0.01069512590765953, + 0.013712920248508453, + 0.004978482145816088, + 0.011353306472301483, + 0.012110048905014992, + 0.012027432210743427, + 0.00537971081212163, + 0.013116875663399696, + 0.02850816585123539, + 0.009568766690790653, + 0.004481087438762188, + 0.12014610320329666, + 0.09636227786540985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.007624808698892593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style consistent with classic 1980s anime featuring bold black outlines and saturated colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera perspective shaky, unstable, and handheld from a first-person view on a ship deck?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive wooden galleon visible in the distance with cannons firing orange and yellow flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a black cannonball streak across the sky and crash into the water causing a massive eruption of sea spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene filled with thick white and gray smoke, gunpowder haze, and multiple explosions on the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635303281247616 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.004635303281247616, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635303281247616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046722494065761566, + 0.012917880900204182, + 0.02722504921257496, + 0.03819848969578743, + 0.22016501426696777, + 0.1207645907998085, + 0.08819998800754547, + 0.022647568956017494, + 0.21967920660972595, + 0.20347975194454193 + ], + "credit_map": [ + 0.006295935716480017, + 0.009291510097682476, + 0.005114700645208359, + 0.004786019679158926, + 0.005368257872760296, + 0.005848133936524391, + 0.005511611234396696, + 0.004506322089582682, + 0.0016432710690423846, + 0.0019268408650532365, + 0.0012475301045924425, + 0.0011443523690104485, + 0.002000784734264016, + 0.0016193961491808295, + 0.0016447752714157104, + 0.0016909298719838262, + 0.0032478231005370617, + 0.005091920495033264, + 0.003206246066838503, + 0.003900009673088789, + 0.0038350282702594995, + 0.0032617393881082535, + 0.00299497926607728, + 0.0016872986452654004, + 0.002807440236210823, + 0.003468806389719248, + 0.006922482512891293, + 0.00595021853223443, + 0.003725164569914341, + 0.002603229833766818, + 0.00714605487883091, + 0.0055750892497599125, + 0.02349645085632801, + 0.008201079443097115, + 0.005721119232475758, + 0.00874093733727932, + 0.053657129406929016, + 0.012688352726399899, + 0.03884235396981239, + 0.06881757080554962, + 0.027576519176363945, + 0.006193890236318111, + 0.008870171383023262, + 0.013704518787562847, + 0.017028450965881348, + 0.015269757248461246, + 0.015635771676898003, + 0.01648550294339657, + 0.00899368803948164, + 0.014267394319176674, + 0.008238366805016994, + 0.014027616940438747, + 0.012140263803303242, + 0.010105679742991924, + 0.012271201238036156, + 0.00815576408058405, + 0.002168426988646388, + 0.004227844532579184, + 0.002518813591450453, + 0.003464330919086933, + 0.0027638347819447517, + 0.0038730159867554903, + 0.0021002416033297777, + 0.0015310613671317697, + 0.03466809168457985, + 0.016228361055254936, + 0.07835681736469269, + 0.02025713212788105, + 0.022013287991285324, + 0.009465972892940044, + 0.020318899303674698, + 0.018370622768998146, + 0.007318009622395039, + 0.003425379516556859, + 0.00782682839781046, + 0.017798984423279762, + 0.0069647980853915215, + 0.0038485857658088207, + 0.08395977318286896, + 0.07233736664056778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.004635303281247616 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_13/index_tensor([5759, 5759])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_13/index_tensor([5759, 5759])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1ed1f36cfe01e93ce5a4bfd2ae9906b9f5a7b872 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_13/index_tensor([5759, 5759])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle drone shot rapidly follows a bright, electric-blue energy spark as it surges through a vast, complex miniature city. The camera tracks the spark's movement, dynamically weaving through a canyon of towering electronic components. The central motif is a bursting sphere of pure plasma, advancing at decisive speed along a path of glowing silicon, forming a complex grid-like street system pulsating with a steady, soft cyan light. This miniature metropolis is densely packed with massive components that dwarf the energy spark: monolithic, matte-black microprocessors stand like black skyscrapers, their surfaces geometrically intricate, while massive, metallic capacitors resemble polished water towers. As the spark travels along its path, its intense, dazzling light casts elongated, rapidly moving shadows, momentarily illuminating the microscopic textures of the silicon and metal surfaces of the components. It's a fluid, high-speed ballet, the spark perfectly navigating sharp turns and paths, leaving faint, ephemeral trails of light in its wake. The scenes are bathed in a cool palette dominated by electric blues and deep, velvety blacks, creating an atmosphere of high-tech, focused, and vibrant raw data processing.", + "seed": 5852, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_13/index_5759_seed_5852_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.055714894901029766, + "vlm_logprob": -0.055714894901029766, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.055714894901029766, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003744021989405155 + ], + "opposite_logprob_tokens": [ + -5.878744125366211 + ], + "logprob_reward": -0.003744021989405155, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.003744021989405155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0997500866651535, + 0.14217454195022583, + 0.08103340864181519, + 0.09346818923950195, + 0.20490574836730957, + 0.06811509281396866, + 0.0631975308060646, + 0.07585974037647247, + 0.0911368653178215, + 0.0803588256239891 + ], + "credit_map": [ + 0.010766712948679924, + 0.00940769538283348, + 0.011860184371471405, + 0.023698840290308, + 0.01941482536494732, + 0.007748064119368792, + 0.009428811259567738, + 0.007424951530992985, + 0.014151358976960182, + 0.01025408785790205, + 0.009929925203323364, + 0.01534135639667511, + 0.015812888741493225, + 0.0403265655040741, + 0.015070201829075813, + 0.021288150921463966, + 0.010317454114556313, + 0.016930267214775085, + 0.010080515407025814, + 0.006932005751878023, + 0.009903634898364544, + 0.012162729166448116, + 0.007945342920720577, + 0.006761458702385426, + 0.011436494067311287, + 0.014693436212837696, + 0.016307834535837173, + 0.008528686128556728, + 0.012807820923626423, + 0.012370175682008266, + 0.008715190924704075, + 0.00860854983329773, + 0.019457979127764702, + 0.03135304152965546, + 0.01648044027388096, + 0.01783892884850502, + 0.04811718314886093, + 0.02552514150738716, + 0.0253971628844738, + 0.020735865458846092, + 0.009268070571124554, + 0.019050180912017822, + 0.009964457713067532, + 0.006332659162580967, + 0.006519338581711054, + 0.005200498271733522, + 0.004881891887634993, + 0.006897995714098215, + 0.006456795614212751, + 0.011531852185726166, + 0.008815454319119453, + 0.00548598263412714, + 0.011352257803082466, + 0.005945650860667229, + 0.007731452584266663, + 0.00587807921692729, + 0.00980196800082922, + 0.009891188703477383, + 0.012272653169929981, + 0.0061075263656675816, + 0.01724192313849926, + 0.008249218575656414, + 0.006011581048369408, + 0.006283679977059364, + 0.010239239782094955, + 0.009739046916365623, + 0.011362534947693348, + 0.008004375733435154, + 0.006764001213014126, + 0.008152002468705177, + 0.021277032792568207, + 0.015598638914525509, + 0.00895757507532835, + 0.014514511451125145, + 0.008821472525596619, + 0.008710566908121109, + 0.016437062993645668, + 0.0069582038559019566, + 0.007589528802782297, + 0.008369906805455685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.003744021989405155 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520921230316162 + ], + "opposite_logprob_tokens": [ + -1.5020921230316162 + ], + "logprob_reward": -0.2520921230316162, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520921230316162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2632763981819153, + 0.1185537576675415, + 0.08617629110813141, + 0.05273794010281563, + 0.09087948501110077, + 0.1740148812532425, + 0.10529915988445282, + 0.0448988676071167, + 0.032446108758449554, + 0.03171703219413757 + ], + "credit_map": [ + 0.03470191732048988, + 0.022478004917502403, + 0.026554197072982788, + 0.06530315428972244, + 0.05705854669213295, + 0.023121999576687813, + 0.017486143857240677, + 0.01657247729599476, + 0.019050396978855133, + 0.009311213158071041, + 0.011891011148691177, + 0.012236511334776878, + 0.025277849286794662, + 0.01807771809399128, + 0.010377646423876286, + 0.012331428937613964, + 0.014879313297569752, + 0.013231683522462845, + 0.011018287390470505, + 0.00918885413557291, + 0.007756880950182676, + 0.013241229578852654, + 0.0064015258103609085, + 0.010458524338901043, + 0.004678219091147184, + 0.00900813564658165, + 0.00749235088005662, + 0.006394547875970602, + 0.008803579956293106, + 0.006453121080994606, + 0.005752608645707369, + 0.004155377857387066, + 0.011380455456674099, + 0.01070951297879219, + 0.007959431037306786, + 0.00698365131393075, + 0.02268262393772602, + 0.011179562658071518, + 0.010226869955658913, + 0.0097573883831501, + 0.04531634971499443, + 0.017199572175741196, + 0.03407526761293411, + 0.012728887610137463, + 0.022422030568122864, + 0.017675820738077164, + 0.010954529978334904, + 0.013642449863255024, + 0.015698134899139404, + 0.011155295185744762, + 0.020207352936267853, + 0.008505316451191902, + 0.012179624289274216, + 0.011345772072672844, + 0.014878308400511742, + 0.011329367756843567, + 0.007035174407064915, + 0.005615920294076204, + 0.006921113468706608, + 0.00505857914686203, + 0.004318222403526306, + 0.004525963682681322, + 0.005829759873449802, + 0.005594139918684959, + 0.003353773383423686, + 0.0028246459551155567, + 0.004509366117417812, + 0.0024496119003742933, + 0.002296621212735772, + 0.0027228808030486107, + 0.008773940615355968, + 0.005515273194760084, + 0.003640310140326619, + 0.004297672305256128, + 0.004328964278101921, + 0.002996599767357111, + 0.003993603866547346, + 0.0025666041765362024, + 0.0037994906306266785, + 0.006093789357692003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.2520921230316162 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007190502947196364 + ], + "opposite_logprob_tokens": [ + -7.25071907043457 + ], + "logprob_reward": -0.0007190502947196364, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007190502947196364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3274327516555786, + 0.037575703114271164, + 0.13018152117729187, + 0.04068134352564812, + 0.04098443314433098, + 0.043606583029031754, + 0.08902134746313095, + 0.08608386665582657, + 0.11845409125089645, + 0.08597837388515472 + ], + "credit_map": [ + 0.04105623811483383, + 0.02563384175300598, + 0.03550514578819275, + 0.0911765918135643, + 0.07734793424606323, + 0.021295934915542603, + 0.020473945885896683, + 0.014943104237318039, + 0.005344514735043049, + 0.003492330899462104, + 0.003432708792388439, + 0.003905676305294037, + 0.005470676347613335, + 0.006906886585056782, + 0.004493599757552147, + 0.004529308527708054, + 0.022034287452697754, + 0.02583865076303482, + 0.01748623140156269, + 0.01298900693655014, + 0.011487217620015144, + 0.020457997918128967, + 0.009781586937606335, + 0.010106552392244339, + 0.00373697723262012, + 0.00542838778346777, + 0.007440485525876284, + 0.00408231932669878, + 0.00666381511837244, + 0.004961037077009678, + 0.004319299012422562, + 0.00404902221634984, + 0.004898251965641975, + 0.005306420847773552, + 0.0046097696758806705, + 0.004889261908829212, + 0.008609863929450512, + 0.00510965334251523, + 0.003762831212952733, + 0.0037983781658113003, + 0.0048169842921197414, + 0.01082178857177496, + 0.005046871956437826, + 0.00446888105943799, + 0.005355296656489372, + 0.005626424215734005, + 0.004065266344696283, + 0.0034050708636641502, + 0.014586610719561577, + 0.011614646762609482, + 0.017234712839126587, + 0.01182206068187952, + 0.008373833261430264, + 0.007051660213619471, + 0.009638556279242039, + 0.00869926530867815, + 0.012396655045449734, + 0.01025690883398056, + 0.014034077525138855, + 0.007637606933712959, + 0.018471337854862213, + 0.009543209336698055, + 0.007101956754922867, + 0.006642114371061325, + 0.011843808926641941, + 0.010259270668029785, + 0.016208596527576447, + 0.007869474589824677, + 0.008444290608167648, + 0.008224992081522942, + 0.03398856520652771, + 0.021615084260702133, + 0.011401994153857231, + 0.010945962741971016, + 0.010602311231195927, + 0.010551140643656254, + 0.010745355859398842, + 0.008932194672524929, + 0.009708265773952007, + 0.01309114694595337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0007190502947196364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072005033493042 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.016072005033493042, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072005033493042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12222066521644592, + 0.06589563190937042, + 0.17722560465335846, + 0.0650656521320343, + 0.11979737132787704, + 0.12542931735515594, + 0.05851750448346138, + 0.11954990029335022, + 0.12175499647855759, + 0.02454333007335663 + ], + "credit_map": [ + 0.012243920005857944, + 0.010958833619952202, + 0.012876716442406178, + 0.04231076315045357, + 0.014254062436521053, + 0.00990194734185934, + 0.011030389927327633, + 0.00864403136074543, + 0.011037060990929604, + 0.006362404674291611, + 0.006699276622384787, + 0.006762978155165911, + 0.009566244669258595, + 0.011147980578243732, + 0.006517364643514156, + 0.007802322506904602, + 0.022634750232100487, + 0.045112308114767075, + 0.021078888326883316, + 0.013449625112116337, + 0.014795596711337566, + 0.02944490872323513, + 0.014741492457687855, + 0.01596803590655327, + 0.00685103889554739, + 0.01423767302185297, + 0.00645252363756299, + 0.010148313827812672, + 0.008361274376511574, + 0.008563559502363205, + 0.00518385274335742, + 0.005267419386655092, + 0.014041218906641006, + 0.017733171582221985, + 0.01439992431551218, + 0.008528534322977066, + 0.02428453601896763, + 0.011961303651332855, + 0.014933351427316666, + 0.01391532551497221, + 0.019118204712867737, + 0.027974793687462807, + 0.022381365299224854, + 0.01020803116261959, + 0.01606687158346176, + 0.012872735969722271, + 0.008897473104298115, + 0.007909850217401981, + 0.00745511194691062, + 0.007874882780015469, + 0.011623747646808624, + 0.008114134892821312, + 0.006605187430977821, + 0.00471778167411685, + 0.006790307816118002, + 0.005336350295692682, + 0.019759397953748703, + 0.01731177791953087, + 0.019607562571763992, + 0.012240838259458542, + 0.01729719527065754, + 0.01149683352559805, + 0.012271115556359291, + 0.009565179236233234, + 0.016911480575799942, + 0.01376664824783802, + 0.015193741768598557, + 0.013536493293941021, + 0.010815110988914967, + 0.013986427336931229, + 0.021578991785645485, + 0.015966106206178665, + 0.0033028051257133484, + 0.0028932748828083277, + 0.003094003302976489, + 0.0031899414025247097, + 0.002617649268358946, + 0.002961099613457918, + 0.002741865813732147, + 0.0037426918279379606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.016072005033493042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059472741559147835 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.0059472741559147835, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059472741559147835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12899698317050934, + 0.05551576614379883, + 0.06819804012775421, + 0.07540017366409302, + 0.06040120869874954, + 0.0661080852150917, + 0.038252219557762146, + 0.1010759100317955, + 0.3383677899837494, + 0.06768383085727692 + ], + "credit_map": [ + 0.014355920255184174, + 0.014631603844463825, + 0.014268280006945133, + 0.01698273979127407, + 0.022106312215328217, + 0.021865908056497574, + 0.012127524241805077, + 0.012658687308430672, + 0.005892434157431126, + 0.004413573071360588, + 0.005623334087431431, + 0.007644610479474068, + 0.007618080358952284, + 0.011372566223144531, + 0.004645314067602158, + 0.008305845782160759, + 0.008605890907347202, + 0.012805846519768238, + 0.010452957823872566, + 0.006448111496865749, + 0.007287956774234772, + 0.009815296158194542, + 0.005040887277573347, + 0.00774108711630106, + 0.009085486643016338, + 0.011949101462960243, + 0.008638588711619377, + 0.006543808151036501, + 0.01243064645677805, + 0.00951602216809988, + 0.010383513756096363, + 0.006853002589195967, + 0.008678433485329151, + 0.010369352996349335, + 0.007102672941982746, + 0.006188253406435251, + 0.008524302393198013, + 0.006628109607845545, + 0.008171399123966694, + 0.004738682880997658, + 0.010137468576431274, + 0.012065628543496132, + 0.007951433770358562, + 0.006477024406194687, + 0.011270876973867416, + 0.0064509036019444466, + 0.0052421605214476585, + 0.006512583699077368, + 0.004346978850662708, + 0.0051864031702280045, + 0.00598307466134429, + 0.005411509890109301, + 0.0034517222084105015, + 0.003226508852094412, + 0.006169626489281654, + 0.0044763945043087006, + 0.013418948277831078, + 0.015975458547472954, + 0.012914680875837803, + 0.007915844209492207, + 0.021014180034399033, + 0.010799810290336609, + 0.008030949160456657, + 0.011006029322743416, + 0.033372387290000916, + 0.033126119524240494, + 0.04424121603369713, + 0.026990823447704315, + 0.026820901781320572, + 0.026808781549334526, + 0.09104158729314804, + 0.05596597120165825, + 0.007191584445536137, + 0.012527468614280224, + 0.007381416391581297, + 0.008761256001889706, + 0.012236906215548515, + 0.006711840629577637, + 0.0061740209348499775, + 0.006699333433061838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0059472741559147835 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone shot rapidly follows a bright, electric-blue energy spark as it surges through a vast, complex miniature city. The camera tracks the spark's movement, dynamically weaving through a canyon of towering electronic components. The central motif is a bursting sphere of pure plasma, advancing at decisive speed along a path of glowing silicon, forming a complex grid-like street system pulsating with a steady, soft cyan light. This miniature metropolis is densely packed with massive components that dwarf the energy spark: monolithic, matte-black microprocessors stand like black skyscrapers, their surfaces geometrically intricate, while massive, metallic capacitors resemble polished water towers. As the spark travels along its path, its intense, dazzling light casts elongated, rapidly moving shadows, momentarily illuminating the microscopic textures of the silicon and metal surfaces of the components. It's a fluid, high-speed ballet, the spark perfectly navigating sharp turns and paths, leaving faint, ephemeral trails of light in its wake. The scenes are bathed in a cool palette dominated by electric blues and deep, velvety blacks, creating an atmosphere of high-tech, focused, and vibrant raw data processing.", + "seed": 5853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_13/index_5759_seed_5853_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07558913449756802, + "vlm_logprob": -0.07558913449756802, + "vlm_yes_no_margin": 4.574999988079071, + "vlm_token_credit": -0.07558913449756802, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0033767367713153362 + ], + "opposite_logprob_tokens": [ + -6.1283769607543945 + ], + "logprob_reward": -0.0033767367713153362, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0033767367713153362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16068850457668304, + 0.19056947529315948, + 0.08034592866897583, + 0.299739271402359, + 0.061010099947452545, + 0.04611481353640556, + 0.04236333817243576, + 0.06316247582435608, + 0.02685091458261013, + 0.029155120253562927 + ], + "credit_map": [ + 0.018172306939959526, + 0.01501030009239912, + 0.04323107749223709, + 0.016241300851106644, + 0.011425669305026531, + 0.03312872350215912, + 0.01169458869844675, + 0.011784546077251434, + 0.016446026042103767, + 0.010859726928174496, + 0.051947660744190216, + 0.05501128360629082, + 0.013034258037805557, + 0.016120921820402145, + 0.013744490221142769, + 0.01340512465685606, + 0.009301133453845978, + 0.010691218078136444, + 0.011725191958248615, + 0.013977638445794582, + 0.008989864960312843, + 0.011612207628786564, + 0.006534257438033819, + 0.007514426484704018, + 0.011769368313252926, + 0.1050402820110321, + 0.011797087267041206, + 0.09497518837451935, + 0.014064936898648739, + 0.01799672842025757, + 0.0161853339523077, + 0.027910325676202774, + 0.006416699383407831, + 0.006058149039745331, + 0.009039627388119698, + 0.00625299708917737, + 0.009114575572311878, + 0.0064295995980501175, + 0.007405017036944628, + 0.010293434374034405, + 0.0033167360816150904, + 0.0057369195856153965, + 0.008796516805887222, + 0.004439906682819128, + 0.00528387725353241, + 0.004112252499908209, + 0.004114438779652119, + 0.010314167477190495, + 0.006346256006509066, + 0.004734129644930363, + 0.0039506941102445126, + 0.005433442071080208, + 0.006604329217225313, + 0.0048302882350981236, + 0.00574088841676712, + 0.004723310470581055, + 0.016179274767637253, + 0.010473915375769138, + 0.009713350795209408, + 0.009915309958159924, + 0.003959254827350378, + 0.004210614133626223, + 0.0035958653315901756, + 0.00511489063501358, + 0.003893892979249358, + 0.0053092436864972115, + 0.00459206523373723, + 0.005777059122920036, + 0.0014128532493487, + 0.002700396114960313, + 0.0012135995784774423, + 0.001951804501004517, + 0.0034193817991763353, + 0.003390740603208542, + 0.004986668936908245, + 0.007177615072578192, + 0.0022574355825781822, + 0.00255973800085485, + 0.002223211107775569, + 0.0031403303146362305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.0033767367713153362 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486087918281555 + ], + "opposite_logprob_tokens": [ + -1.2236087322235107 + ], + "logprob_reward": -0.3486087918281555, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3486087918281555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1229998916387558, + 0.06731602549552917, + 0.09335380047559738, + 0.10866393893957138, + 0.032911669462919235, + 0.048806801438331604, + 0.09946690499782562, + 0.23372520506381989, + 0.03546131029725075, + 0.15729448199272156 + ], + "credit_map": [ + 0.021153975278139114, + 0.018422558903694153, + 0.017339175567030907, + 0.020255442708730698, + 0.008556392043828964, + 0.01817767135798931, + 0.007439361419528723, + 0.011655318550765514, + 0.010563814081251621, + 0.0059637511149048805, + 0.011726239696145058, + 0.010909894481301308, + 0.006344016641378403, + 0.007861301302909851, + 0.00645610922947526, + 0.007490898948162794, + 0.00916225928813219, + 0.014235510490834713, + 0.015065856277942657, + 0.013041822239756584, + 0.008348644711077213, + 0.01753675565123558, + 0.006260427180677652, + 0.009702527895569801, + 0.0048780255019664764, + 0.0352809876203537, + 0.003917142748832703, + 0.03182953968644142, + 0.007489567156881094, + 0.007602334953844547, + 0.007193088997155428, + 0.010473255068063736, + 0.005409049801528454, + 0.003954352345317602, + 0.003738456405699253, + 0.00408001197502017, + 0.004012570716440678, + 0.0031837227288633585, + 0.004475957714021206, + 0.00405754754319787, + 0.004259314388036728, + 0.008548839017748833, + 0.007203894667327404, + 0.006450338289141655, + 0.005151865538209677, + 0.006172132212668657, + 0.0038231294602155685, + 0.007197288796305656, + 0.007909934036433697, + 0.014588477090001106, + 0.010537565685808659, + 0.017780909314751625, + 0.015438877046108246, + 0.01020023226737976, + 0.010856325738132, + 0.012154584750533104, + 0.041446033865213394, + 0.021219586953520775, + 0.023189645260572433, + 0.06617854535579681, + 0.00986573938280344, + 0.044404760003089905, + 0.011196869425475597, + 0.016224026679992676, + 0.0035509685985744, + 0.007855716161429882, + 0.003935134503990412, + 0.007313860114663839, + 0.003647840116173029, + 0.00320771848782897, + 0.0032094772905111313, + 0.0027405961882323027, + 0.021978089585900307, + 0.017311155796051025, + 0.024212948977947235, + 0.03008623793721199, + 0.010286913253366947, + 0.018545541912317276, + 0.014829453080892563, + 0.02004413865506649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.3486087918281555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010417040903121233 + ], + "opposite_logprob_tokens": [ + -6.876041889190674 + ], + "logprob_reward": -0.0010417040903121233, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010417040903121233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08156093209981918, + 0.13080087304115295, + 0.11020568758249283, + 0.426822304725647, + 0.059315528720617294, + 0.042225394397974014, + 0.0438675694167614, + 0.05260046198964119, + 0.026027632877230644, + 0.02657361701130867 + ], + "credit_map": [ + 0.009660798124969006, + 0.010838394984602928, + 0.013089083135128021, + 0.01364162564277649, + 0.00760147999972105, + 0.010234033688902855, + 0.008163698948919773, + 0.008331825956702232, + 0.011810925789177418, + 0.009529158473014832, + 0.034043435007333755, + 0.03405088186264038, + 0.008784535340964794, + 0.01363956369459629, + 0.008178828284144402, + 0.010763541795313358, + 0.013075043447315693, + 0.014955897815525532, + 0.023468218743801117, + 0.020339839160442352, + 0.008922663517296314, + 0.014931011945009232, + 0.007424527779221535, + 0.007088479585945606, + 0.019952572882175446, + 0.1510128527879715, + 0.019364850595593452, + 0.1258406788110733, + 0.01907767355442047, + 0.028188809752464294, + 0.023402005434036255, + 0.03998285531997681, + 0.0072214617393910885, + 0.007372856605798006, + 0.009883109480142593, + 0.007193517405539751, + 0.004544581286609173, + 0.00647388631477952, + 0.005750744603574276, + 0.010875371284782887, + 0.0029311119578778744, + 0.009799227118492126, + 0.010478214360773563, + 0.003391215344890952, + 0.004005718976259232, + 0.00327604659833014, + 0.0021656008902937174, + 0.0061782593838870525, + 0.004323222208768129, + 0.010008451528847218, + 0.0035352404229342937, + 0.008776630274951458, + 0.004849126562476158, + 0.004027726128697395, + 0.002951620612293482, + 0.005395547952502966, + 0.007932168431580067, + 0.00485930684953928, + 0.0050032916478812695, + 0.01725209690630436, + 0.0029636435210704803, + 0.006805833429098129, + 0.003637849586084485, + 0.004146269056946039, + 0.002580177504569292, + 0.005781290587037802, + 0.0035244992468506098, + 0.005168958101421595, + 0.00240500969812274, + 0.0022657485678792, + 0.0018833390204235911, + 0.0024186100345104933, + 0.003249066649004817, + 0.0027297709602862597, + 0.004424747079610825, + 0.005906565114855766, + 0.001750558614730835, + 0.002768797567114234, + 0.0022569610737264156, + 0.003487149253487587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.0010417040903121233 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184026703238487 + ], + "opposite_logprob_tokens": [ + -4.018184185028076 + ], + "logprob_reward": -0.018184026703238487, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184026703238487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1556706577539444, + 0.0759596899151802, + 0.08589672297239304, + 0.19025278091430664, + 0.10179489105939865, + 0.09018569439649582, + 0.10453592985868454, + 0.15169793367385864, + 0.02180088870227337, + 0.022204827517271042 + ], + "credit_map": [ + 0.02092145010828972, + 0.021841444075107574, + 0.03316514939069748, + 0.01807938702404499, + 0.009286998771131039, + 0.025013098493218422, + 0.009920552372932434, + 0.017442572861909866, + 0.007370509207248688, + 0.004919370636343956, + 0.017039863392710686, + 0.018263237550854683, + 0.008025040850043297, + 0.005524814128875732, + 0.006851916201412678, + 0.007964935153722763, + 0.008576322346925735, + 0.010755762457847595, + 0.010859041474759579, + 0.010242887772619724, + 0.011987348087131977, + 0.019007496535778046, + 0.006657568737864494, + 0.007810293696820736, + 0.008590101264417171, + 0.06206778436899185, + 0.008757119998335838, + 0.0561399981379509, + 0.0101376473903656, + 0.012697839178144932, + 0.01387329213321209, + 0.017989005893468857, + 0.01552459504455328, + 0.012889092788100243, + 0.015507562085986137, + 0.009873207658529282, + 0.007893306203186512, + 0.010108708404004574, + 0.011692914180457592, + 0.01830550655722618, + 0.00764794135466218, + 0.012917025946080685, + 0.016622595489025116, + 0.008548695594072342, + 0.009819754399359226, + 0.013169815763831139, + 0.007235881872475147, + 0.014223986305296421, + 0.011922094970941544, + 0.015299972146749496, + 0.011226503178477287, + 0.021744241937994957, + 0.011418172158300877, + 0.01116762775927782, + 0.007845081388950348, + 0.01391222607344389, + 0.023843510076403618, + 0.017899755388498306, + 0.007804390974342823, + 0.055410388857126236, + 0.012245038524270058, + 0.01063372939825058, + 0.01103280484676361, + 0.012828304432332516, + 0.0025536674074828625, + 0.004662917926907539, + 0.003002483630552888, + 0.0036344334948807955, + 0.001861365744844079, + 0.0021368972957134247, + 0.0022496110759675503, + 0.0016995128244161606, + 0.0021261172369122505, + 0.0026547443121671677, + 0.003458599792793393, + 0.004532285500317812, + 0.002458244329318404, + 0.002282467670738697, + 0.0023367400281131268, + 0.0023556288797408342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.018184026703238487 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734413094818592 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734413094818592, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734413094818592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17631612718105316, + 0.07713411748409271, + 0.13052807748317719, + 0.08055153489112854, + 0.19199106097221375, + 0.14666765928268433, + 0.05209803208708763, + 0.05808699131011963, + 0.04217303544282913, + 0.044453348964452744 + ], + "credit_map": [ + 0.02526724338531494, + 0.015368946827948093, + 0.03345765918493271, + 0.022405797615647316, + 0.019799992442131042, + 0.032547976821660995, + 0.01435396820306778, + 0.01311453990638256, + 0.0072427974082529545, + 0.005245414096862078, + 0.019607791677117348, + 0.018649626523256302, + 0.006231752689927816, + 0.006899661850184202, + 0.006095249205827713, + 0.007161828223615885, + 0.018211238086223602, + 0.015707502141594887, + 0.013933456502854824, + 0.015907494351267815, + 0.014333691447973251, + 0.025740962475538254, + 0.00839658547192812, + 0.018297135829925537, + 0.009462879039347172, + 0.011573223397135735, + 0.009217510931193829, + 0.013199873268604279, + 0.0077918223105371, + 0.010105928406119347, + 0.008716439828276634, + 0.010483860969543457, + 0.026056010276079178, + 0.03070526383817196, + 0.02323741465806961, + 0.022763757035136223, + 0.024120774120092392, + 0.01690404862165451, + 0.022633112967014313, + 0.025570673868060112, + 0.009837067686021328, + 0.043072544038295746, + 0.017669014632701874, + 0.012774499133229256, + 0.014140457846224308, + 0.016990676522254944, + 0.009497364051640034, + 0.022686025127768517, + 0.005226318724453449, + 0.006608250085264444, + 0.005802186205983162, + 0.007887424901127815, + 0.007734348997473717, + 0.006258275825530291, + 0.005318074021488428, + 0.007263153325766325, + 0.011972802691161633, + 0.009440547786653042, + 0.010760839097201824, + 0.00510592944920063, + 0.003847802057862282, + 0.008739791810512543, + 0.0037024274934083223, + 0.004516851156949997, + 0.005640065297484398, + 0.00844748318195343, + 0.007724561262875795, + 0.008377835154533386, + 0.002437221584841609, + 0.00392756424844265, + 0.0029000258073210716, + 0.0027182805351912975, + 0.004502867814153433, + 0.006118536926805973, + 0.006220641545951366, + 0.008935138583183289, + 0.005262545309960842, + 0.004599269945174456, + 0.004373599775135517, + 0.004440748132765293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.006734413094818592 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone shot rapidly follows a bright, electric-blue energy spark as it surges through a vast, complex miniature city. The camera tracks the spark's movement, dynamically weaving through a canyon of towering electronic components. The central motif is a bursting sphere of pure plasma, advancing at decisive speed along a path of glowing silicon, forming a complex grid-like street system pulsating with a steady, soft cyan light. This miniature metropolis is densely packed with massive components that dwarf the energy spark: monolithic, matte-black microprocessors stand like black skyscrapers, their surfaces geometrically intricate, while massive, metallic capacitors resemble polished water towers. As the spark travels along its path, its intense, dazzling light casts elongated, rapidly moving shadows, momentarily illuminating the microscopic textures of the silicon and metal surfaces of the components. It's a fluid, high-speed ballet, the spark perfectly navigating sharp turns and paths, leaving faint, ephemeral trails of light in its wake. The scenes are bathed in a cool palette dominated by electric blues and deep, velvety blacks, creating an atmosphere of high-tech, focused, and vibrant raw data processing.", + "seed": 5854, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_13/index_5759_seed_5854_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0652583490125835, + "vlm_logprob": -0.0652583490125835, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.0652583490125835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025830259546637535 + ], + "opposite_logprob_tokens": [ + -6.127583026885986 + ], + "logprob_reward": -0.0025830259546637535, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0025830259546637535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11095431447029114, + 0.47881296277046204, + 0.02051570452749729, + 0.0644417405128479, + 0.07190801203250885, + 0.07979582995176315, + 0.04124259203672409, + 0.06652823090553284, + 0.027024850249290466, + 0.038775697350502014 + ], + "credit_map": [ + 0.018687307834625244, + 0.021846506744623184, + 0.014051713980734348, + 0.011585358530282974, + 0.01424605306237936, + 0.010220871306955814, + 0.013287094421684742, + 0.007029424421489239, + 0.07857757061719894, + 0.042377449572086334, + 0.03296530246734619, + 0.05942004919052124, + 0.06361772119998932, + 0.04209635779261589, + 0.04272501915693283, + 0.11703352630138397, + 0.002759727416560054, + 0.0021241751965135336, + 0.003109120298177004, + 0.0029435751494020224, + 0.0019233273342251778, + 0.003850771114230156, + 0.0020892166066914797, + 0.0017157936235889792, + 0.006261331029236317, + 0.007175097241997719, + 0.007792254444211721, + 0.00922176893800497, + 0.006960313301533461, + 0.009049997664988041, + 0.012329662218689919, + 0.005651318933814764, + 0.01229689922183752, + 0.00873644184321165, + 0.008232212625443935, + 0.009661278687417507, + 0.006216586101800203, + 0.01255748514086008, + 0.004409248009324074, + 0.009797876700758934, + 0.00617609079927206, + 0.024931415915489197, + 0.019943971186876297, + 0.004727323539555073, + 0.006410945672541857, + 0.0051244874484837055, + 0.006927798967808485, + 0.00555380480363965, + 0.00373471318744123, + 0.005148252937942743, + 0.004158928990364075, + 0.012804617173969746, + 0.0036920648999512196, + 0.005052817985415459, + 0.0033286376856267452, + 0.0033225668594241142, + 0.007977881468832493, + 0.010884488932788372, + 0.00610967306420207, + 0.009671377018094063, + 0.009677150286734104, + 0.009551705792546272, + 0.006472291424870491, + 0.006183679681271315, + 0.0031534316949546337, + 0.002188484650105238, + 0.0030881420243531466, + 0.004252900369465351, + 0.0032329352106899023, + 0.0032141609117388725, + 0.003704063594341278, + 0.004190734587609768, + 0.004967020824551582, + 0.006161767989397049, + 0.003968675155192614, + 0.00449965987354517, + 0.0040383185259997845, + 0.004442539997398853, + 0.006918619386851788, + 0.003779099090024829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0025830259546637535 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31343889236450195 + ], + "opposite_logprob_tokens": [ + -1.313438892364502 + ], + "logprob_reward": -0.31343889236450195, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31343889236450195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11001510173082352, + 0.4054212272167206, + 0.11903470754623413, + 0.020073194056749344, + 0.033412300050258636, + 0.13367755711078644, + 0.1036708801984787, + 0.047215402126312256, + 0.013277593068778515, + 0.014201992191374302 + ], + "credit_map": [ + 0.014675783924758434, + 0.014205838553607464, + 0.01394026167690754, + 0.01592155732214451, + 0.01634177751839161, + 0.009783775545656681, + 0.015696130692958832, + 0.009449979290366173, + 0.04766986891627312, + 0.026211515069007874, + 0.03337721526622772, + 0.10622884333133698, + 0.043142277747392654, + 0.03145069628953934, + 0.028949949890375137, + 0.08839086443185806, + 0.009881462901830673, + 0.008822734467685223, + 0.007357132621109486, + 0.010784962214529514, + 0.037532106041908264, + 0.014052142389118671, + 0.00944050308316946, + 0.021163662895560265, + 0.0020119412802159786, + 0.002064487664029002, + 0.002801739377900958, + 0.003378470428287983, + 0.002437175717204809, + 0.0030925602186471224, + 0.0021473760716617107, + 0.0021394439972937107, + 0.006336890626698732, + 0.0039002022240310907, + 0.0033301380462944508, + 0.0035973258782178164, + 0.002912292256951332, + 0.005392786581069231, + 0.0030395991634577513, + 0.004903066903352737, + 0.0168303269892931, + 0.03625691682100296, + 0.031695012003183365, + 0.007787421345710754, + 0.013519901782274246, + 0.011663641780614853, + 0.007438269443809986, + 0.008486058562994003, + 0.020794006064534187, + 0.007455804385244846, + 0.00620185025036335, + 0.044159259647130966, + 0.007110539358109236, + 0.006762966513633728, + 0.0070030223578214645, + 0.0041834344156086445, + 0.005344647914171219, + 0.00706550432369113, + 0.005574130918830633, + 0.005412400234490633, + 0.008196178823709488, + 0.007670399267226458, + 0.0036472005303949118, + 0.004304940812289715, + 0.0015939364675432444, + 0.0012275944463908672, + 0.0017490409081801772, + 0.0018930830992758274, + 0.0019308192422613502, + 0.0016282832948490977, + 0.0016324887983500957, + 0.0016223471611738205, + 0.001971959602087736, + 0.0025046139489859343, + 0.0018233363516628742, + 0.0014257486909627914, + 0.0020072488114237785, + 0.0020008129067718983, + 0.0014076465740799904, + 0.001060625072568655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.31343889236450195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00104301399551332 + ], + "opposite_logprob_tokens": [ + -6.87604284286499 + ], + "logprob_reward": -0.00104301399551332, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.00104301399551332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09293260425329208, + 0.33684706687927246, + 0.11278388649225235, + 0.05059325322508812, + 0.04367614910006523, + 0.053315356373786926, + 0.11511967331171036, + 0.10098635405302048, + 0.028070785105228424, + 0.06567493826150894 + ], + "credit_map": [ + 0.01333948876708746, + 0.00984933227300644, + 0.012798638083040714, + 0.012096836231648922, + 0.01500686164945364, + 0.00982253160327673, + 0.010842912830412388, + 0.009176001884043217, + 0.047612980008125305, + 0.02733091078698635, + 0.026150129735469818, + 0.12888428568840027, + 0.02600123919546604, + 0.0305434949696064, + 0.02613101340830326, + 0.024192966520786285, + 0.011683136224746704, + 0.007899352349340916, + 0.007635493762791157, + 0.011669393628835678, + 0.02892751805484295, + 0.013567507266998291, + 0.011186448857188225, + 0.02021504007279873, + 0.0050201863050460815, + 0.0048556458204984665, + 0.004424515645951033, + 0.0073562078177928925, + 0.004311548080295324, + 0.007914508692920208, + 0.012568810023367405, + 0.004141824319958687, + 0.006456997245550156, + 0.006141459569334984, + 0.0060780588537454605, + 0.0056978254579007626, + 0.005023374687880278, + 0.00484187388792634, + 0.003340167226269841, + 0.006096390541642904, + 0.006441781762987375, + 0.01139145903289318, + 0.010967453010380268, + 0.004313524346798658, + 0.005394273903220892, + 0.005813692696392536, + 0.0045427982695400715, + 0.004450376611202955, + 0.02474275231361389, + 0.011922510340809822, + 0.009028367698192596, + 0.03460438922047615, + 0.0073698293417692184, + 0.007613812107592821, + 0.009348570369184017, + 0.010489443317055702, + 0.014138534665107727, + 0.016951559111475945, + 0.012512383982539177, + 0.009661090560257435, + 0.011319594457745552, + 0.014344247989356518, + 0.010517788119614124, + 0.011541156098246574, + 0.0032408167608082294, + 0.005670817568898201, + 0.0028796307742595673, + 0.0036029317416250706, + 0.0031085938680917025, + 0.0033772431779652834, + 0.0027755973860621452, + 0.0034151547588407993, + 0.006557728163897991, + 0.0055900029838085175, + 0.01870514452457428, + 0.003663813229650259, + 0.004663215018808842, + 0.020559193566441536, + 0.0036246925592422485, + 0.002311148215085268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.00104301399551332 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738912779837847 + ], + "opposite_logprob_tokens": [ + -5.006739139556885 + ], + "logprob_reward": -0.006738912779837847, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738912779837847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09642468392848969, + 0.35014796257019043, + 0.12438897043466568, + 0.09224803000688553, + 0.09875626862049103, + 0.05509732663631439, + 0.0629645511507988, + 0.05807995796203613, + 0.01924302615225315, + 0.04264921694993973 + ], + "credit_map": [ + 0.015541817992925644, + 0.016592545434832573, + 0.012013796716928482, + 0.011793367564678192, + 0.007990650832653046, + 0.009663389064371586, + 0.015311836265027523, + 0.007517282385379076, + 0.06583218276500702, + 0.024348169565200806, + 0.02773863635957241, + 0.03973502293229103, + 0.03923935070633888, + 0.03176466003060341, + 0.027996722608804703, + 0.09349322319030762, + 0.0084675969555974, + 0.008441525511443615, + 0.0074200378730893135, + 0.01042743120342493, + 0.04042264446616173, + 0.012407440692186356, + 0.010207682847976685, + 0.026594607159495354, + 0.008570907637476921, + 0.006176639348268509, + 0.009368103928864002, + 0.012974939309060574, + 0.008995907381176949, + 0.013227584771811962, + 0.02590499073266983, + 0.007028953172266483, + 0.016189854592084885, + 0.011048320680856705, + 0.010600238107144833, + 0.012889696285128593, + 0.007733995094895363, + 0.020482569932937622, + 0.006337813567370176, + 0.013473776169121265, + 0.007911152206361294, + 0.010816482827067375, + 0.009757829830050468, + 0.004391475114971399, + 0.00556932482868433, + 0.0070581333711743355, + 0.004304354544728994, + 0.005288572516292334, + 0.007312724832445383, + 0.006588185206055641, + 0.008047809824347496, + 0.02116057090461254, + 0.00489585567265749, + 0.00545445317402482, + 0.004654646851122379, + 0.004850301891565323, + 0.0069929370656609535, + 0.007715890649706125, + 0.008153104223310947, + 0.00892715621739626, + 0.0066000367514789104, + 0.005268053617328405, + 0.007049353793263435, + 0.007373427040874958, + 0.002612510696053505, + 0.0023999835830181837, + 0.002303800545632839, + 0.002480023307725787, + 0.00189143605530262, + 0.0028752826619893312, + 0.0022965455427765846, + 0.002383443294093013, + 0.005035150330513716, + 0.004342169966548681, + 0.006056601647287607, + 0.005381464026868343, + 0.004729833919554949, + 0.007319713942706585, + 0.004623336251825094, + 0.005160948261618614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.006738912779837847 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024878999684005976 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.0024878999684005976, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024878999684005976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07323002815246582, + 0.24982210993766785, + 0.08651647716760635, + 0.10613998025655746, + 0.12246651202440262, + 0.10865859687328339, + 0.045013219118118286, + 0.05418572202324867, + 0.03791278973221779, + 0.11605449765920639 + ], + "credit_map": [ + 0.01122164074331522, + 0.013778163120150566, + 0.009291768074035645, + 0.007368308957666159, + 0.008784252218902111, + 0.005699330009520054, + 0.010901405476033688, + 0.006185167469084263, + 0.04353782907128334, + 0.021350987255573273, + 0.019316408783197403, + 0.07647886872291565, + 0.022519893944263458, + 0.022576147690415382, + 0.021909426897764206, + 0.022132564336061478, + 0.007907281629741192, + 0.0064445617608726025, + 0.005090482998639345, + 0.005907769780606031, + 0.02992568537592888, + 0.009534568525850773, + 0.007284428924322128, + 0.014421705156564713, + 0.009074158035218716, + 0.01083177886903286, + 0.013819782063364983, + 0.01947193406522274, + 0.010828525759279728, + 0.015302288345992565, + 0.018372420221567154, + 0.008439101278781891, + 0.017047667875885963, + 0.014280060306191444, + 0.016309725120663643, + 0.014550000429153442, + 0.011046621948480606, + 0.022253001108765602, + 0.008136515505611897, + 0.01884293369948864, + 0.01007099449634552, + 0.04007860645651817, + 0.0351916179060936, + 0.0037948100361973047, + 0.0063669378869235516, + 0.004609289579093456, + 0.0045519364066421986, + 0.003994408529251814, + 0.008470876142382622, + 0.004672615323215723, + 0.0043358937837183475, + 0.014191905036568642, + 0.002959831152111292, + 0.004977992735803127, + 0.002934439107775688, + 0.002469669096171856, + 0.006428144406527281, + 0.008361941203474998, + 0.005709097255021334, + 0.009026273153722286, + 0.006461534649133682, + 0.007614727597683668, + 0.0050038788467645645, + 0.005580132827162743, + 0.004028687719255686, + 0.005742963403463364, + 0.004731402732431889, + 0.005302611738443375, + 0.0038833876606076956, + 0.003949357196688652, + 0.005077283829450607, + 0.0051971017383039, + 0.009989558719098568, + 0.007659386843442917, + 0.016152115538716316, + 0.019501222297549248, + 0.011513721197843552, + 0.019519787281751633, + 0.01628127135336399, + 0.01543744932860136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.0024878999684005976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone shot rapidly follows a bright, electric-blue energy spark as it surges through a vast, complex miniature city. The camera tracks the spark's movement, dynamically weaving through a canyon of towering electronic components. The central motif is a bursting sphere of pure plasma, advancing at decisive speed along a path of glowing silicon, forming a complex grid-like street system pulsating with a steady, soft cyan light. This miniature metropolis is densely packed with massive components that dwarf the energy spark: monolithic, matte-black microprocessors stand like black skyscrapers, their surfaces geometrically intricate, while massive, metallic capacitors resemble polished water towers. As the spark travels along its path, its intense, dazzling light casts elongated, rapidly moving shadows, momentarily illuminating the microscopic textures of the silicon and metal surfaces of the components. It's a fluid, high-speed ballet, the spark perfectly navigating sharp turns and paths, leaving faint, ephemeral trails of light in its wake. The scenes are bathed in a cool palette dominated by electric blues and deep, velvety blacks, creating an atmosphere of high-tech, focused, and vibrant raw data processing.", + "seed": 5855, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_13/index_5759_seed_5855_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06687805003020912, + "vlm_logprob": -0.06687805003020912, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.06687805003020912, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002500266768038273 + ], + "opposite_logprob_tokens": [ + -6.252500057220459 + ], + "logprob_reward": -0.002500266768038273, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.002500266768038273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2876175045967102, + 0.058283984661102295, + 0.021129276603460312, + 0.3919568657875061, + 0.05050661042332649, + 0.022648783400654793, + 0.017933785915374756, + 0.05426699295639992, + 0.024849222972989082, + 0.07080692797899246 + ], + "credit_map": [ + 0.0459531806409359, + 0.033395182341337204, + 0.04935452342033386, + 0.029712270945310593, + 0.01896740309894085, + 0.07154975831508636, + 0.01756906695663929, + 0.021116113290190697, + 0.009745440445840359, + 0.00540304696187377, + 0.01302989199757576, + 0.005098475608974695, + 0.007182464003562927, + 0.005110412836074829, + 0.006263640243560076, + 0.006450612563639879, + 0.00238418229855597, + 0.002164687030017376, + 0.002663726219907403, + 0.0025924211367964745, + 0.003076611552387476, + 0.003717539133504033, + 0.002446535276249051, + 0.0020835748873651028, + 0.1556023806333542, + 0.01748359203338623, + 0.13220107555389404, + 0.02542312629520893, + 0.015812285244464874, + 0.015532033517956734, + 0.012804295867681503, + 0.017098067328333855, + 0.00588175468146801, + 0.006560737732797861, + 0.005540968384593725, + 0.005256363656371832, + 0.0067373644560575485, + 0.00953160785138607, + 0.006327588111162186, + 0.004670223221182823, + 0.002073215786367655, + 0.002437639283016324, + 0.0026822793297469616, + 0.002198276109993458, + 0.003773593110963702, + 0.0029027119744569063, + 0.0027961262967437506, + 0.003784940578043461, + 0.002278056461364031, + 0.0043062251061201096, + 0.0022307599429041147, + 0.001786727923899889, + 0.002317708218470216, + 0.002123782876878977, + 0.0013802131870761514, + 0.001510311965830624, + 0.006477923132479191, + 0.00690237432718277, + 0.006352559197694063, + 0.005242995452135801, + 0.009636232629418373, + 0.008023638278245926, + 0.0056049455888569355, + 0.006026324816048145, + 0.002440133597701788, + 0.004247406497597694, + 0.001887911930680275, + 0.0038944664411246777, + 0.003513394622132182, + 0.0031978688202798367, + 0.0028200680390000343, + 0.0028479734901338816, + 0.00960193295031786, + 0.013299024663865566, + 0.012694272212684155, + 0.012208407744765282, + 0.006290554068982601, + 0.005467899609357119, + 0.004817349370568991, + 0.006427485030144453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.002500266768038273 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31343957781791687 + ], + "opposite_logprob_tokens": [ + -1.3134396076202393 + ], + "logprob_reward": -0.31343957781791687, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31343957781791687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1523304581642151, + 0.06944286078214645, + 0.09409335255622864, + 0.17896124720573425, + 0.06793263554573059, + 0.08068931847810745, + 0.04090439900755882, + 0.16649268567562103, + 0.04589444398880005, + 0.1032586321234703 + ], + "credit_map": [ + 0.022526821121573448, + 0.012130646966397762, + 0.02611105889081955, + 0.028419379144906998, + 0.015105321072041988, + 0.016628099605441093, + 0.013845673762261868, + 0.017563464120030403, + 0.013005631975829601, + 0.009145335294306278, + 0.016062214970588684, + 0.006669939495623112, + 0.006139177363365889, + 0.005343498196452856, + 0.0061693452298641205, + 0.006907715927809477, + 0.012993989512324333, + 0.011712977662682533, + 0.014749323017895222, + 0.009228132665157318, + 0.013449892401695251, + 0.01052086241543293, + 0.010434606112539768, + 0.01100356224924326, + 0.07412029802799225, + 0.009105379693210125, + 0.06284018605947495, + 0.012062783353030682, + 0.004741115495562553, + 0.005579845979809761, + 0.004654051270335913, + 0.005857575684785843, + 0.010740567930042744, + 0.012273267842829227, + 0.008806388825178146, + 0.00868228916078806, + 0.006803001742810011, + 0.008020023815333843, + 0.005995230283588171, + 0.006611858960241079, + 0.008197554387152195, + 0.00706423120573163, + 0.013388953171670437, + 0.006135626696050167, + 0.007491162046790123, + 0.010947085916996002, + 0.011057998985052109, + 0.016406705603003502, + 0.005297050811350346, + 0.007109958212822676, + 0.005947527475655079, + 0.005183687433600426, + 0.004929553251713514, + 0.004423449281603098, + 0.003980046138167381, + 0.004033125005662441, + 0.023248925805091858, + 0.02223804220557213, + 0.02645810693502426, + 0.02077958919107914, + 0.01893634721636772, + 0.02364255115389824, + 0.014164399355649948, + 0.017024703323841095, + 0.004168936517089605, + 0.006922343745827675, + 0.003945603966712952, + 0.005774280522018671, + 0.009506182745099068, + 0.005498081911355257, + 0.0052157677710056305, + 0.004863245412707329, + 0.015142559073865414, + 0.01045900583267212, + 0.014014916494488716, + 0.014907903037965298, + 0.01139771193265915, + 0.011427952907979488, + 0.01119176484644413, + 0.014716815203428268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.31343957781791687 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010421803453937173 + ], + "opposite_logprob_tokens": [ + -6.876042366027832 + ], + "logprob_reward": -0.0010421803453937173, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010421803453937173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29358866810798645, + 0.0827934667468071, + 0.06856128573417664, + 0.30917641520500183, + 0.03665357083082199, + 0.018027782440185547, + 0.02246450074017048, + 0.049696363508701324, + 0.03719070181250572, + 0.08184725046157837 + ], + "credit_map": [ + 0.05578653886914253, + 0.031845834106206894, + 0.04020237922668457, + 0.08611439913511276, + 0.02272018976509571, + 0.021985100582242012, + 0.01763862557709217, + 0.017295576632022858, + 0.010184737853705883, + 0.011747096665203571, + 0.010135268792510033, + 0.015468087047338486, + 0.008142728358507156, + 0.008910602889955044, + 0.006156422197818756, + 0.012048522010445595, + 0.008878130465745926, + 0.007210997864603996, + 0.010652055963873863, + 0.008525272831320763, + 0.013893414288759232, + 0.008224619552493095, + 0.0060080657713115215, + 0.005168725270777941, + 0.12731385231018066, + 0.01594630256295204, + 0.10919582098722458, + 0.02085595764219761, + 0.008160527795553207, + 0.008761451579630375, + 0.00846441276371479, + 0.010478074662387371, + 0.0039381603710353374, + 0.005153109319508076, + 0.005709959659725428, + 0.005356962326914072, + 0.0045285336673259735, + 0.004627184476703405, + 0.003999911714345217, + 0.0033397474326193333, + 0.0019488242687657475, + 0.0021333463955670595, + 0.002421446843072772, + 0.0017759131733328104, + 0.002107870066538453, + 0.0028513132128864527, + 0.00212606112472713, + 0.0026630063075572252, + 0.0035937351640313864, + 0.004832013510167599, + 0.00263070035725832, + 0.0023227501660585403, + 0.002873246790841222, + 0.002603588392958045, + 0.0013742460869252682, + 0.002234221203252673, + 0.007698071654886007, + 0.00828683003783226, + 0.00531522324308753, + 0.006151772569864988, + 0.007386763114482164, + 0.006194171961396933, + 0.004534537438303232, + 0.00412899162620306, + 0.0030014761723577976, + 0.007784142158925533, + 0.0030773039907217026, + 0.004097995813935995, + 0.007502521853893995, + 0.00359865534119308, + 0.004273303784430027, + 0.0038553045596927404, + 0.009447121061384678, + 0.010039610788226128, + 0.012164001353085041, + 0.014743213541805744, + 0.009815413504838943, + 0.01098401565104723, + 0.006155194714665413, + 0.00849868543446064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0010421803453937173 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978246983140707 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.00978246983140707, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978246983140707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12259067595005035, + 0.10344008356332779, + 0.04011949896812439, + 0.4070344567298889, + 0.0417710579931736, + 0.04372958466410637, + 0.06145220994949341, + 0.09484981745481491, + 0.023593921214342117, + 0.06141870841383934 + ], + "credit_map": [ + 0.0176690686494112, + 0.017443208023905754, + 0.015975434333086014, + 0.020851977169513702, + 0.014574475586414337, + 0.01640162244439125, + 0.009689737111330032, + 0.00998515635728836, + 0.012898148968815804, + 0.011243738234043121, + 0.01959754340350628, + 0.010786314494907856, + 0.01440225075930357, + 0.009358533658087254, + 0.010263646021485329, + 0.014889910817146301, + 0.005163851194083691, + 0.004545411095023155, + 0.0046247499994933605, + 0.0032731087412685156, + 0.009353024885058403, + 0.006311874836683273, + 0.003607080550864339, + 0.0032404002267867327, + 0.17160452902317047, + 0.018822044134140015, + 0.1473555564880371, + 0.027182908728718758, + 0.009850805625319481, + 0.011575200594961643, + 0.00787025410681963, + 0.012773159891366959, + 0.00466478057205677, + 0.004933732096105814, + 0.006301974877715111, + 0.003871734021231532, + 0.005417265463620424, + 0.007638719864189625, + 0.004895111080259085, + 0.004047741647809744, + 0.004244114272296429, + 0.005176721140742302, + 0.005851646885275841, + 0.005322623532265425, + 0.006416282616555691, + 0.007532461080700159, + 0.004965991247445345, + 0.004219743423163891, + 0.00650379853323102, + 0.010094987228512764, + 0.005794794298708439, + 0.006268731784075499, + 0.01389939896762371, + 0.007556228898465633, + 0.006420502904802561, + 0.004913768731057644, + 0.013586126267910004, + 0.01553034596145153, + 0.014483529143035412, + 0.011751215904951096, + 0.010940495878458023, + 0.011182406917214394, + 0.009103290736675262, + 0.008272411301732063, + 0.002260050969198346, + 0.004200638737529516, + 0.0028441352769732475, + 0.003275094320997596, + 0.0033209947869181633, + 0.0025310141500085592, + 0.0027649488765746355, + 0.0023970440961420536, + 0.0076032415963709354, + 0.005441777873784304, + 0.017267365008592606, + 0.0054970127530395985, + 0.007154365070164204, + 0.006068710703402758, + 0.005924382247030735, + 0.006461854558438063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.00978246983140707 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bright, electric-blue energy spark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned as a high-angle drone shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a miniature city made of towering electronic components like microprocessors and capacitors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a steady, soft cyan light forming a grid-like street system?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spark leave faint, ephemeral trails of light as it moves rapidly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762575538828969 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.00762575538828969, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762575538828969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12972316145896912, + 0.07118501514196396, + 0.04274175316095352, + 0.48610320687294006, + 0.06890202313661575, + 0.03708002343773842, + 0.02979114279150963, + 0.05723775178194046, + 0.029319696128368378, + 0.047916196286678314 + ], + "credit_map": [ + 0.013174746185541153, + 0.011265506036579609, + 0.016055919229984283, + 0.025499440729618073, + 0.010116674937307835, + 0.037640031427145004, + 0.008162854239344597, + 0.007807992864400148, + 0.007048761006444693, + 0.007868092507123947, + 0.008699893951416016, + 0.007685151416808367, + 0.012662539258599281, + 0.008929146453738213, + 0.009676272049546242, + 0.0086151622235775, + 0.005104042589664459, + 0.005440152250230312, + 0.006036676932126284, + 0.003931308165192604, + 0.00776751060038805, + 0.006620359607040882, + 0.0042297751642763615, + 0.003611927619203925, + 0.19725258648395538, + 0.022115271538496017, + 0.17009107768535614, + 0.031770672649145126, + 0.016005510464310646, + 0.016760429367423058, + 0.014762742444872856, + 0.01734490692615509, + 0.006108887959271669, + 0.007231026887893677, + 0.00880265049636364, + 0.00729357311502099, + 0.008073176257312298, + 0.014838733710348606, + 0.008667055517435074, + 0.0078869154676795, + 0.003441621083766222, + 0.0039930157363414764, + 0.00469689816236496, + 0.003576449351385236, + 0.003986332565546036, + 0.0055289361625909805, + 0.00491249468177557, + 0.0069442749954760075, + 0.0035252308007329702, + 0.0050004166550934315, + 0.0022343387827277184, + 0.0029380377382040024, + 0.007178154774010181, + 0.003856796771287918, + 0.002497076289728284, + 0.0025610881857573986, + 0.007337621413171291, + 0.010252130217850208, + 0.006920323241502047, + 0.007397449109703302, + 0.0074120922945439816, + 0.007934153079986572, + 0.0053011872805655, + 0.004682791884988546, + 0.0026151537895202637, + 0.009427567012608051, + 0.0031311451457440853, + 0.002125572180375457, + 0.003687004093080759, + 0.002613592427223921, + 0.0017595432000234723, + 0.003960116766393185, + 0.005324999801814556, + 0.007416195701807737, + 0.00837900023907423, + 0.009575391188263893, + 0.004331096541136503, + 0.004416246898472309, + 0.0041170865297317505, + 0.004356182646006346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.00762575538828969 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_14/index_tensor([8761, 8761])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_14/index_tensor([8761, 8761])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..de0356993733af9db7adddc1dbc70db327e05062 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_14/index_tensor([8761, 8761])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a high-contrast digital look, the slow motion of a 360-degree orbiting camera captures a moment of intense repulsion, frozen in extreme slow motion within a desolate black void. At the center of the frame, a heavy, solid block of black ink and a delicate white porcelain pen holder are forcefully flung away from each other after an unseen, powerful collision. The inkstone, a rectangular block composed of matte, dark slate, slowly rotates along its axis, its dense shape remaining largely intact, save for a small, newly chiseled corner that gleams in the harsh light. In contrast, the cylindrical porcelain pen holder shatters catastrophically, its thin, lustrous white fragments ejected outward in an expanding sphere. A beautiful, chaotic cloud of fine white ceramic dust erupts from the point of impact, catching the single, powerful keylight and emitting an ethereal glow, while sharper, grainy black particles from the ink shoot outward from this cloud in more direct trajectories, like microscopic obsidian fragments. As the camera gracefully orbits this levitating explosion, high-contrast lighting sculpts out every stunning detail—the frantic spiderweb textures on shattered porcelain, the smooth, cool surface of ink, and individual specks of dust—creating a scene of controlled chaos of breathtaking beauty and technical precision.", + "seed": 8026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_14/index_8761_seed_8026_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0224863076582551, + "vlm_logprob": -0.0224863076582551, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.0224863076582551, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017390504479408264 + ], + "opposite_logprob_tokens": [ + -4.267390727996826 + ], + "logprob_reward": -0.017390504479408264, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017390504479408264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12746533751487732, + 0.05491361767053604, + 0.04072548821568489, + 0.09475786983966827, + 0.04123976081609726, + 0.07945980876684189, + 0.1544226109981537, + 0.044940438121557236, + 0.09685197472572327, + 0.26522302627563477 + ], + "credit_map": [ + 0.023649703711271286, + 0.0113245053216815, + 0.01567111536860466, + 0.007920530624687672, + 0.008742387406527996, + 0.024409687146544456, + 0.0070889731869101524, + 0.028658436611294746, + 0.006218509282916784, + 0.010634016245603561, + 0.007259341888129711, + 0.00642778817564249, + 0.004879602696746588, + 0.008819986134767532, + 0.00274355779401958, + 0.007930816151201725, + 0.0065589603036642075, + 0.0027554568368941545, + 0.008262359537184238, + 0.003139703068882227, + 0.002155308611690998, + 0.0072937533259391785, + 0.003059348091483116, + 0.007500596344470978, + 0.01721741445362568, + 0.003936417866498232, + 0.018394291400909424, + 0.0058776442892849445, + 0.0028107089456170797, + 0.018042834475636482, + 0.005169183015823364, + 0.023309364914894104, + 0.006444170139729977, + 0.0023990704212337732, + 0.01022409088909626, + 0.003352268598973751, + 0.0014752807328477502, + 0.005593856330960989, + 0.004173105116933584, + 0.007577915675938129, + 0.01539595890790224, + 0.0032600522972643375, + 0.014432062394917011, + 0.005399769172072411, + 0.002717193216085434, + 0.012264441698789597, + 0.006035321857780218, + 0.01995501108467579, + 0.028398538008332253, + 0.006294631864875555, + 0.04281654953956604, + 0.00825869757682085, + 0.005304013378918171, + 0.02385525219142437, + 0.008790583349764347, + 0.030704351142048836, + 0.005648260936141014, + 0.0029233035165816545, + 0.011686138808727264, + 0.003818181110545993, + 0.0022687383461743593, + 0.006203344091773033, + 0.004595700651407242, + 0.007796773687005043, + 0.014987914822995663, + 0.004352007061243057, + 0.01833033375442028, + 0.007704558316618204, + 0.0033413292840123177, + 0.01389024406671524, + 0.009458716958761215, + 0.024786872789263725, + 0.053455449640750885, + 0.013753693550825119, + 0.06361488997936249, + 0.017311761155724525, + 0.01108294166624546, + 0.03555409982800484, + 0.020139893516898155, + 0.050310298800468445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.017390504479408264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867476224899292 + ], + "opposite_logprob_tokens": [ + -3.0486748218536377 + ], + "logprob_reward": -0.04867476224899292, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867476224899292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11315914988517761, + 0.07033205032348633, + 0.07200266420841217, + 0.052169688045978546, + 0.06292802095413208, + 0.0346575565636158, + 0.08707311749458313, + 0.14938890933990479, + 0.06625693291425705, + 0.2920319736003876 + ], + "credit_map": [ + 0.01841127686202526, + 0.01997458189725876, + 0.014500883407890797, + 0.008461542427539825, + 0.011050996370613575, + 0.017347168177366257, + 0.005744271446019411, + 0.017668426036834717, + 0.0078024291433393955, + 0.009946450591087341, + 0.008284091018140316, + 0.006698443554341793, + 0.007458265870809555, + 0.013754607178270817, + 0.004422089550644159, + 0.0119656752794981, + 0.014556484296917915, + 0.005334688350558281, + 0.010986845009028912, + 0.005457459948956966, + 0.004761707503348589, + 0.012477309443056583, + 0.005728420335799456, + 0.012699754908680916, + 0.006818836089223623, + 0.005155527498573065, + 0.008785190992057323, + 0.007564328145235777, + 0.004096615128219128, + 0.00736516248434782, + 0.005728846415877342, + 0.006655182223767042, + 0.008739881217479706, + 0.004740634933114052, + 0.013894585892558098, + 0.007590945810079575, + 0.003974286373704672, + 0.009468658827245235, + 0.005998113192617893, + 0.008520917035639286, + 0.005122879520058632, + 0.002563049551099539, + 0.005029269028455019, + 0.0036751755978912115, + 0.002627449342980981, + 0.004499365575611591, + 0.0036478734109550714, + 0.007492495235055685, + 0.012032209895551205, + 0.007048458326607943, + 0.011297882534563541, + 0.018026860430836678, + 0.00620291568338871, + 0.01119319349527359, + 0.011345839127898216, + 0.009925756603479385, + 0.021609755232930183, + 0.009140124544501305, + 0.033120863139629364, + 0.013089319691061974, + 0.007889009080827236, + 0.02119729109108448, + 0.0176277793943882, + 0.025714769959449768, + 0.009765864349901676, + 0.005352570675313473, + 0.015051990747451782, + 0.005630661733448505, + 0.003484271001070738, + 0.009562880732119083, + 0.00572203891351819, + 0.011686655692756176, + 0.0552799291908741, + 0.015448862686753273, + 0.061393316835165024, + 0.023661233484745026, + 0.01521548256278038, + 0.04058819264173508, + 0.02627674676477909, + 0.0541682094335556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.04867476224899292 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004650728777050972 + ], + "opposite_logprob_tokens": [ + -5.379650592803955 + ], + "logprob_reward": -0.004650728777050972, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004650728777050972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17917101085186005, + 0.14855435490608215, + 0.05370647832751274, + 0.08626821637153625, + 0.05466708168387413, + 0.05077217519283295, + 0.06957092136144638, + 0.09675589948892593, + 0.11646807193756104, + 0.144065722823143 + ], + "credit_map": [ + 0.03548792377114296, + 0.01682962477207184, + 0.022825581952929497, + 0.010520635172724724, + 0.011617654003202915, + 0.03759083151817322, + 0.007196586579084396, + 0.037102192640304565, + 0.01775551214814186, + 0.03227832168340683, + 0.018177591264247894, + 0.01348774041980505, + 0.011235477402806282, + 0.02538323402404785, + 0.005883653648197651, + 0.02435283362865448, + 0.010369195602834225, + 0.004456439521163702, + 0.008233851753175259, + 0.0051497300155460835, + 0.003460949519649148, + 0.008092278614640236, + 0.004437289200723171, + 0.00950674433261156, + 0.012513586319983006, + 0.005990156438201666, + 0.02031315304338932, + 0.007208194583654404, + 0.004630817100405693, + 0.013279036618769169, + 0.006960300263017416, + 0.015372984111309052, + 0.008710294030606747, + 0.004078300669789314, + 0.008010883815586567, + 0.006097572855651379, + 0.00342820817604661, + 0.009558557532727718, + 0.006298528052866459, + 0.008484740741550922, + 0.00834651943296194, + 0.0027150141540914774, + 0.012576338835060596, + 0.005245239473879337, + 0.0023890314623713493, + 0.007372153922915459, + 0.004399377852678299, + 0.007728505879640579, + 0.009586451575160027, + 0.0049415514804422855, + 0.0145711125805974, + 0.011127657257020473, + 0.004561906680464745, + 0.008652967400848866, + 0.007127524819225073, + 0.009001757018268108, + 0.011884527280926704, + 0.00862767081707716, + 0.01837572641670704, + 0.013942119665443897, + 0.0059078107587993145, + 0.012774062342941761, + 0.010862389579415321, + 0.01438160240650177, + 0.016095928847789764, + 0.0108492998406291, + 0.014849727042019367, + 0.010734917595982552, + 0.007674726657569408, + 0.014272257685661316, + 0.013877762481570244, + 0.028113463893532753, + 0.02951725944876671, + 0.007252625189721584, + 0.03448051959276199, + 0.010118965990841389, + 0.006062009837478399, + 0.019200051203370094, + 0.01255571935325861, + 0.02487858198583126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.004650728777050972 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809729963541031 + ], + "opposite_logprob_tokens": [ + -3.288097381591797 + ], + "logprob_reward": -0.03809729963541031, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809729963541031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12442910671234131, + 0.13331647217273712, + 0.06716007739305496, + 0.09420289099216461, + 0.08487527072429657, + 0.17218030989170074, + 0.0642235279083252, + 0.07263042777776718, + 0.07909418642520905, + 0.10788771510124207 + ], + "credit_map": [ + 0.01844402775168419, + 0.016178840771317482, + 0.013531822711229324, + 0.010699503123760223, + 0.008163148537278175, + 0.023529013618826866, + 0.007134952116757631, + 0.026747798547148705, + 0.016473805531859398, + 0.022276658564805984, + 0.014900295995175838, + 0.010435307398438454, + 0.010765764862298965, + 0.025964539498090744, + 0.007137710694223642, + 0.025362396612763405, + 0.010490640066564083, + 0.005375879816710949, + 0.010073043406009674, + 0.005849324632436037, + 0.004561539273709059, + 0.013260981999337673, + 0.004630421753972769, + 0.012918241322040558, + 0.014710220508277416, + 0.0075333514250814915, + 0.018609948456287384, + 0.012943387031555176, + 0.004787161946296692, + 0.013867830857634544, + 0.007211063522845507, + 0.01453991886228323, + 0.015860775485634804, + 0.0062513770535588264, + 0.013832859694957733, + 0.009459340944886208, + 0.00659966329112649, + 0.012547909282147884, + 0.008361920714378357, + 0.011961426585912704, + 0.028499683365225792, + 0.00915021263062954, + 0.03788335248827934, + 0.01879831776022911, + 0.008758875541388988, + 0.02249360829591751, + 0.020474670454859734, + 0.026121584698557854, + 0.010196911171078682, + 0.004152537323534489, + 0.01768341101706028, + 0.006313922815024853, + 0.003148154355585575, + 0.010277258232235909, + 0.0041769822128117085, + 0.008274349384009838, + 0.009134414605796337, + 0.004008392337709665, + 0.01877317763864994, + 0.005945928394794464, + 0.002994103590026498, + 0.010126490145921707, + 0.006830313242971897, + 0.014817608520388603, + 0.01189349964261055, + 0.006036230828613043, + 0.014037368819117546, + 0.011396675370633602, + 0.005008940119296312, + 0.01096974965184927, + 0.01003937516361475, + 0.009712346829473972, + 0.017258182168006897, + 0.00902851764112711, + 0.01934501715004444, + 0.009234074503183365, + 0.00819777324795723, + 0.016411326825618744, + 0.012791966088116169, + 0.015620856545865536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.03809729963541031 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036182431504130363 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.0036182431504130363, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036182431504130363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11762973666191101, + 0.2775225043296814, + 0.10246895253658295, + 0.03879820555448532, + 0.057980384677648544, + 0.07693200558423996, + 0.11803077906370163, + 0.030801640823483467, + 0.05932887643575668, + 0.1205068901181221 + ], + "credit_map": [ + 0.01949787139892578, + 0.017925569787621498, + 0.024062372744083405, + 0.009376592002809048, + 0.008363459259271622, + 0.016120200976729393, + 0.006322390865534544, + 0.01596127264201641, + 0.03677524998784065, + 0.04537444934248924, + 0.033411432057619095, + 0.016683068126440048, + 0.016598539426922798, + 0.060753416270017624, + 0.009690850041806698, + 0.05823551118373871, + 0.01589926704764366, + 0.009736109524965286, + 0.019280899316072464, + 0.010075450874865055, + 0.008280757814645767, + 0.015437362715601921, + 0.008995836600661278, + 0.014763274230062962, + 0.007151240482926369, + 0.0026881746016442776, + 0.008184862323105335, + 0.0050519309006631374, + 0.002172076376155019, + 0.004693756811320782, + 0.004099625628441572, + 0.004756540060043335, + 0.00869155116379261, + 0.0028503448702394962, + 0.011670645326375961, + 0.007310999091714621, + 0.002573873847723007, + 0.010430142283439636, + 0.004047554451972246, + 0.010405272245407104, + 0.012826872989535332, + 0.0039006907027214766, + 0.016876371577382088, + 0.0070595815777778625, + 0.004935841076076031, + 0.00958968885242939, + 0.010912607423961163, + 0.010830353014171124, + 0.013833708129823208, + 0.008268829435110092, + 0.023150615394115448, + 0.0240088552236557, + 0.006501609925180674, + 0.014806539751589298, + 0.013069577515125275, + 0.014391041360795498, + 0.004437863826751709, + 0.0027039728593081236, + 0.00617050938308239, + 0.003043142845854163, + 0.0020588755141943693, + 0.003944849595427513, + 0.003314055036753416, + 0.005128371063619852, + 0.011208884418010712, + 0.00407909182831645, + 0.006941104773432016, + 0.0038714995607733727, + 0.0035089896991848946, + 0.007638080511242151, + 0.006082338280975819, + 0.01599888689815998, + 0.021685009822249413, + 0.006251809187233448, + 0.024863632395863533, + 0.00889595691114664, + 0.00582102220505476, + 0.016155391931533813, + 0.010727139189839363, + 0.026106925681233406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.0036182431504130363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-contrast digital look, the slow motion of a 360-degree orbiting camera captures a moment of intense repulsion, frozen in extreme slow motion within a desolate black void. At the center of the frame, a heavy, solid block of black ink and a delicate white porcelain pen holder are forcefully flung away from each other after an unseen, powerful collision. The inkstone, a rectangular block composed of matte, dark slate, slowly rotates along its axis, its dense shape remaining largely intact, save for a small, newly chiseled corner that gleams in the harsh light. In contrast, the cylindrical porcelain pen holder shatters catastrophically, its thin, lustrous white fragments ejected outward in an expanding sphere. A beautiful, chaotic cloud of fine white ceramic dust erupts from the point of impact, catching the single, powerful keylight and emitting an ethereal glow, while sharper, grainy black particles from the ink shoot outward from this cloud in more direct trajectories, like microscopic obsidian fragments. As the camera gracefully orbits this levitating explosion, high-contrast lighting sculpts out every stunning detail—the frantic spiderweb textures on shattered porcelain, the smooth, cool surface of ink, and individual specks of dust—creating a scene of controlled chaos of breathtaking beauty and technical precision.", + "seed": 8027, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_14/index_8761_seed_8027_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04295228431001306, + "vlm_logprob": -0.04295228431001306, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.04295228431001306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007143788039684296 + ], + "opposite_logprob_tokens": [ + -5.382143974304199 + ], + "logprob_reward": -0.007143788039684296, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.007143788039684296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15101084113121033, + 0.1299959123134613, + 0.06885844469070435, + 0.042946044355630875, + 0.1384623646736145, + 0.1035558208823204, + 0.054305609315633774, + 0.03823612257838249, + 0.04081013798713684, + 0.23181867599487305 + ], + "credit_map": [ + 0.014494129456579685, + 0.018061071634292603, + 0.014621488749980927, + 0.029813271015882492, + 0.013721365481615067, + 0.023021645843982697, + 0.014379879459738731, + 0.0228979904204607, + 0.011838564649224281, + 0.009121576324105263, + 0.010833151638507843, + 0.011296533048152924, + 0.006491790525615215, + 0.030434507876634598, + 0.011064247228205204, + 0.03891553729772568, + 0.0088784359395504, + 0.007225785404443741, + 0.009658182971179485, + 0.006428510881960392, + 0.007741822395473719, + 0.010105757042765617, + 0.010307933203876019, + 0.008512013591825962, + 0.005184476263821125, + 0.0039198496378958225, + 0.0027169680688530207, + 0.006398695521056652, + 0.004534198436886072, + 0.006238693371415138, + 0.006989097222685814, + 0.006964065600186586, + 0.02202751860022545, + 0.006390114314854145, + 0.01310321781784296, + 0.01136430911719799, + 0.006856195628643036, + 0.029531607404351234, + 0.01655679941177368, + 0.03263259679079056, + 0.010533791966736317, + 0.006359729915857315, + 0.016805369406938553, + 0.009513428434729576, + 0.007980242371559143, + 0.01826515793800354, + 0.023650895804166794, + 0.010447213426232338, + 0.00721267843618989, + 0.004594468977302313, + 0.0080411396920681, + 0.008847814053297043, + 0.0046807145699858665, + 0.007985047996044159, + 0.006154241040349007, + 0.006789502687752247, + 0.0043458337895572186, + 0.002371281385421753, + 0.005886875092983246, + 0.0040182494558393955, + 0.00391140952706337, + 0.007930748164653778, + 0.004708201624453068, + 0.005063528660684824, + 0.00408351281657815, + 0.0030295203905552626, + 0.005597098730504513, + 0.005374704021960497, + 0.005625167395919561, + 0.006202037446200848, + 0.007179465610533953, + 0.0037186285480856895, + 0.030504919588565826, + 0.013288923539221287, + 0.03082313761115074, + 0.024256840348243713, + 0.029334096238017082, + 0.0458826906979084, + 0.02696584165096283, + 0.03076223097741604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.007143788039684296 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003682106733322 + ], + "opposite_logprob_tokens": [ + -2.6950368881225586 + ], + "logprob_reward": -0.07003682106733322, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003682106733322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055768270045518875, + 0.4074024260044098, + 0.033594194799661636, + 0.19489824771881104, + 0.038789212703704834, + 0.02709263376891613, + 0.04236115887761116, + 0.01705087162554264, + 0.019991829991340637, + 0.16305109858512878 + ], + "credit_map": [ + 0.0068712253123521805, + 0.006458937656134367, + 0.005905854515731335, + 0.007295205723494291, + 0.004755373112857342, + 0.009655929170548916, + 0.006054971367120743, + 0.008770783431828022, + 0.0419231615960598, + 0.02685336023569107, + 0.030842414125800133, + 0.03054673969745636, + 0.01772262714803219, + 0.09718925505876541, + 0.03273414075374603, + 0.12959077954292297, + 0.0034559927880764008, + 0.002978626638650894, + 0.00460833590477705, + 0.003368960227817297, + 0.004568545613437891, + 0.0053318580612540245, + 0.003406636882573366, + 0.005875245202332735, + 0.006318834610283375, + 0.0028087215032428503, + 0.0021295880433171988, + 0.0020721459295600653, + 0.002432760316878557, + 0.08704452216625214, + 0.0030128529760986567, + 0.08907885104417801, + 0.00799292977899313, + 0.0017539630644023418, + 0.0034267576411366463, + 0.0025090191047638655, + 0.0018152145203202963, + 0.011439180001616478, + 0.0032790866680443287, + 0.006573069840669632, + 0.0037804574239999056, + 0.0016424189088866115, + 0.0049952357076108456, + 0.0027163559570908546, + 0.002029821742326021, + 0.004976953379809856, + 0.0037583380471915007, + 0.003193055745214224, + 0.004456776659935713, + 0.003997801337391138, + 0.006766949314624071, + 0.004771380685269833, + 0.004693572409451008, + 0.006219306029379368, + 0.006940128281712532, + 0.004515251610428095, + 0.001655659405514598, + 0.002055517164990306, + 0.002341708168387413, + 0.0020999321714043617, + 0.0019256249070167542, + 0.0020596832036972046, + 0.0034686450380831957, + 0.001444103429093957, + 0.002659435151144862, + 0.0013447956880554557, + 0.00277261296287179, + 0.0025714102666825056, + 0.0024551276583224535, + 0.0027873029466718435, + 0.00362081709317863, + 0.0017803326481953263, + 0.018073761835694313, + 0.009003182873129845, + 0.023480745032429695, + 0.0169619619846344, + 0.019006898626685143, + 0.036224257200956345, + 0.016878798604011536, + 0.023421527817845345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 9 + ], + "selected_reward": -0.07003682106733322 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005273479502648115 + ], + "opposite_logprob_tokens": [ + -5.255273342132568 + ], + "logprob_reward": -0.005273479502648115, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005273479502648115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16417963802814484, + 0.16455326974391937, + 0.09796484559774399, + 0.12305189669132233, + 0.08561479300260544, + 0.06344214081764221, + 0.1703188270330429, + 0.023200729861855507, + 0.02403957024216652, + 0.08363426476716995 + ], + "credit_map": [ + 0.014980422332882881, + 0.018826698884367943, + 0.016020487993955612, + 0.02538665197789669, + 0.0138736916705966, + 0.02829211764037609, + 0.014014558866620064, + 0.03278501704335213, + 0.014539926312863827, + 0.01416124776005745, + 0.017741918563842773, + 0.017330091446638107, + 0.010010063648223877, + 0.03294519707560539, + 0.014296024106442928, + 0.04352881386876106, + 0.011475486680865288, + 0.008795542642474174, + 0.011552585288882256, + 0.008431032299995422, + 0.01234568189829588, + 0.015233472920954227, + 0.014106975868344307, + 0.016024073585867882, + 0.008800226263701916, + 0.008690360002219677, + 0.007122037932276726, + 0.014051469974219799, + 0.012965748086571693, + 0.02791198343038559, + 0.019797828048467636, + 0.023712245747447014, + 0.015882255509495735, + 0.0043147290125489235, + 0.00862992461770773, + 0.005383856128901243, + 0.00391023512929678, + 0.01568373292684555, + 0.01082804799079895, + 0.020982012152671814, + 0.006726087536662817, + 0.005275755189359188, + 0.009878471493721008, + 0.008685187436640263, + 0.0046987044624984264, + 0.010271129198372364, + 0.010165314190089703, + 0.0077414982952177525, + 0.018594732508063316, + 0.016536150127649307, + 0.021801266819238663, + 0.025616643950343132, + 0.019773637875914574, + 0.020671769976615906, + 0.030463464558124542, + 0.01686118356883526, + 0.003235895885154605, + 0.0019706683233380318, + 0.00382679863832891, + 0.0020401254296302795, + 0.0018029703060165048, + 0.004323201719671488, + 0.0026378764305263758, + 0.0033631944097578526, + 0.0029709560330957174, + 0.0019189963350072503, + 0.004134632647037506, + 0.002275364240631461, + 0.0037322421558201313, + 0.0037971229758113623, + 0.0032395583111792803, + 0.0019706995226442814, + 0.009629461914300919, + 0.006813631393015385, + 0.011492710560560226, + 0.012193881906569004, + 0.010492080822587013, + 0.013112276792526245, + 0.011828800663352013, + 0.00807142723351717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.005273479502648115 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704962491989136 + ], + "opposite_logprob_tokens": [ + -2.127049684524536 + ], + "logprob_reward": -0.12704962491989136, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704962491989136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1267150342464447, + 0.10285443812608719, + 0.09179790318012238, + 0.05877283960580826, + 0.07974272966384888, + 0.32475990056991577, + 0.05612312629818916, + 0.03272147476673126, + 0.08177672326564789, + 0.044735804200172424 + ], + "credit_map": [ + 0.011011071503162384, + 0.014784171245992184, + 0.012286055833101273, + 0.022921927273273468, + 0.014559716917574406, + 0.020674332976341248, + 0.013237038627266884, + 0.017240719869732857, + 0.012570474296808243, + 0.010236253961920738, + 0.01208763662725687, + 0.009008950553834438, + 0.006436696741729975, + 0.019482648000121117, + 0.008758588694036007, + 0.024273192510008812, + 0.010746254585683346, + 0.009593128226697445, + 0.006925337016582489, + 0.008074522949755192, + 0.019600996747612953, + 0.011969305574893951, + 0.010187922045588493, + 0.014700435101985931, + 0.005969080608338118, + 0.0037130704149603844, + 0.0034856011625379324, + 0.0053219059482216835, + 0.0056418804451823235, + 0.013971338979899883, + 0.009523624554276466, + 0.011146340519189835, + 0.014200178906321526, + 0.003994561266154051, + 0.008600214496254921, + 0.005431837867945433, + 0.004004330839961767, + 0.016778774559497833, + 0.006544747389853001, + 0.02018808387219906, + 0.032768283039331436, + 0.021871063858270645, + 0.0510043203830719, + 0.03569542616605759, + 0.02603979781270027, + 0.04326722398400307, + 0.07957162708044052, + 0.03454216569662094, + 0.004704621154814959, + 0.004987247753888369, + 0.00765185197815299, + 0.007515573874115944, + 0.0056822472251951694, + 0.010321613401174545, + 0.008146194741129875, + 0.007113778498023748, + 0.0033888937905430794, + 0.004082577768713236, + 0.004567301366478205, + 0.005334372632205486, + 0.003746702568605542, + 0.004130909219384193, + 0.004775729496032, + 0.0026949881576001644, + 0.01132349856197834, + 0.006263772491365671, + 0.012097164988517761, + 0.010519045405089855, + 0.009915826842188835, + 0.014047893695533276, + 0.011728639714419842, + 0.005880881566554308, + 0.0035533870104700327, + 0.0034882233012467623, + 0.006144947838038206, + 0.004658399149775505, + 0.006008984986692667, + 0.009724460542201996, + 0.006051347125321627, + 0.005106054246425629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.12704962491989136 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257708020508289 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257708020508289, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257708020508289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14373846352100372, + 0.40252840518951416, + 0.05077309533953667, + 0.03779124841094017, + 0.07932565361261368, + 0.0670999214053154, + 0.04810202866792679, + 0.03906797617673874, + 0.09782428294420242, + 0.033748891204595566 + ], + "credit_map": [ + 0.010562661103904247, + 0.010285482741892338, + 0.008209999650716782, + 0.018861044198274612, + 0.009969295002520084, + 0.036371760070323944, + 0.010132056660950184, + 0.03934616595506668, + 0.042194247245788574, + 0.027580121532082558, + 0.02993037737905979, + 0.03255443647503853, + 0.01725950464606285, + 0.09400641173124313, + 0.03042738139629364, + 0.12857592105865479, + 0.004762243013828993, + 0.005150261335074902, + 0.004934167955070734, + 0.005811248440295458, + 0.0086507648229599, + 0.006970296148210764, + 0.006370556075125933, + 0.008123556151986122, + 0.0025462729390710592, + 0.001619196031242609, + 0.0012445104075595737, + 0.00157153548207134, + 0.0018795382929965854, + 0.013468348421156406, + 0.0033429565373808146, + 0.012118891812860966, + 0.012128747999668121, + 0.004648218397051096, + 0.007166127674281597, + 0.008222791366279125, + 0.004318804480135441, + 0.013561718165874481, + 0.00983502622693777, + 0.019444216042757034, + 0.0071192411705851555, + 0.003615336026996374, + 0.010837777517735958, + 0.0064511592499911785, + 0.005145890172570944, + 0.012039008550345898, + 0.014135997742414474, + 0.007755512837320566, + 0.005554456263780594, + 0.0038225664757192135, + 0.006095638033002615, + 0.004984579514712095, + 0.004666041117161512, + 0.009228905662894249, + 0.0066642737947404385, + 0.0070855687372386456, + 0.00604707607999444, + 0.002754249144345522, + 0.006876134779304266, + 0.0037747325841337442, + 0.0030301199294626713, + 0.0070207188837230206, + 0.0039960285648703575, + 0.005568915978074074, + 0.011222489178180695, + 0.005519229918718338, + 0.015615074895322323, + 0.00957977119833231, + 0.01101759821176529, + 0.019393520429730415, + 0.015852874144911766, + 0.009623727761209011, + 0.004728007595986128, + 0.003790078219026327, + 0.004335024859756231, + 0.004407862666994333, + 0.0034034629352390766, + 0.004753224551677704, + 0.004746017511934042, + 0.0035852116998285055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.005257708020508289 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-contrast digital look, the slow motion of a 360-degree orbiting camera captures a moment of intense repulsion, frozen in extreme slow motion within a desolate black void. At the center of the frame, a heavy, solid block of black ink and a delicate white porcelain pen holder are forcefully flung away from each other after an unseen, powerful collision. The inkstone, a rectangular block composed of matte, dark slate, slowly rotates along its axis, its dense shape remaining largely intact, save for a small, newly chiseled corner that gleams in the harsh light. In contrast, the cylindrical porcelain pen holder shatters catastrophically, its thin, lustrous white fragments ejected outward in an expanding sphere. A beautiful, chaotic cloud of fine white ceramic dust erupts from the point of impact, catching the single, powerful keylight and emitting an ethereal glow, while sharper, grainy black particles from the ink shoot outward from this cloud in more direct trajectories, like microscopic obsidian fragments. As the camera gracefully orbits this levitating explosion, high-contrast lighting sculpts out every stunning detail—the frantic spiderweb textures on shattered porcelain, the smooth, cool surface of ink, and individual specks of dust—creating a scene of controlled chaos of breathtaking beauty and technical precision.", + "seed": 8028, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_14/index_8761_seed_8028_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06822075862437486, + "vlm_logprob": -0.06822075862437486, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.06822075862437486, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011509448289871216 + ], + "opposite_logprob_tokens": [ + -4.636509418487549 + ], + "logprob_reward": -0.011509448289871216, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011509448289871216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22451716661453247, + 0.07080218940973282, + 0.03774791955947876, + 0.18832314014434814, + 0.1253366619348526, + 0.032914407551288605, + 0.12033676356077194, + 0.12733256816864014, + 0.025743702426552773, + 0.046945516020059586 + ], + "credit_map": [ + 0.01822041906416416, + 0.023070674389600754, + 0.022361954674124718, + 0.014600050635635853, + 0.03158736601471901, + 0.03765344247221947, + 0.025173844769597054, + 0.051849402487277985, + 0.0036237025633454323, + 0.00510815717279911, + 0.008679184131324291, + 0.0037661532405763865, + 0.006416307296603918, + 0.02101137302815914, + 0.00565335713326931, + 0.016543950885534286, + 0.0028734919615089893, + 0.003897745395079255, + 0.0043838429264724255, + 0.0036532741505652666, + 0.006029341835528612, + 0.005774215795099735, + 0.0056618512608110905, + 0.0054741548374295235, + 0.009011453948915005, + 0.006097124423831701, + 0.0055999718606472015, + 0.005976175889372826, + 0.005800921004265547, + 0.0741678923368454, + 0.00823436863720417, + 0.0734352394938469, + 0.0076325214467942715, + 0.007846500724554062, + 0.006872282829135656, + 0.02000955492258072, + 0.011407678946852684, + 0.028138622641563416, + 0.015881553292274475, + 0.02754795178771019, + 0.0060585867613554, + 0.002097636926919222, + 0.0024330655578523874, + 0.0038542647380381823, + 0.002337659941986203, + 0.0058466182090342045, + 0.004440181888639927, + 0.005846393760293722, + 0.017417553812265396, + 0.00913681648671627, + 0.0204336978495121, + 0.010176470503211021, + 0.011796819046139717, + 0.017542071640491486, + 0.01966068707406521, + 0.014172650873661041, + 0.013824189081788063, + 0.012111008167266846, + 0.022078143432736397, + 0.008916453458368778, + 0.028293784707784653, + 0.017411114647984505, + 0.010629608295857906, + 0.014068267308175564, + 0.0045550731010735035, + 0.0024641973432153463, + 0.0023518765810877085, + 0.0035320287570357323, + 0.0036922430153936148, + 0.0038383726496249437, + 0.002371113281697035, + 0.002938798163086176, + 0.011661493219435215, + 0.0027244642842561007, + 0.005306662991642952, + 0.005516336299479008, + 0.003095941850915551, + 0.006903086323291063, + 0.0034032987896353006, + 0.00833422876894474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.011509448289871216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004838436841965 + ], + "opposite_logprob_tokens": [ + -2.6950483322143555 + ], + "logprob_reward": -0.07004838436841965, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004838436841965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1294846087694168, + 0.11517415940761566, + 0.08942079544067383, + 0.21884560585021973, + 0.12163019925355911, + 0.02857496589422226, + 0.039789799600839615, + 0.10372458398342133, + 0.06868961453437805, + 0.0846656933426857 + ], + "credit_map": [ + 0.011015152558684349, + 0.014913201332092285, + 0.02330099791288376, + 0.01481037400662899, + 0.017542440444231033, + 0.019757727161049843, + 0.011556566692888737, + 0.01658814586699009, + 0.005132905673235655, + 0.010603649541735649, + 0.010914430022239685, + 0.008980782702565193, + 0.021462174132466316, + 0.026128970086574554, + 0.012496340088546276, + 0.019454898312687874, + 0.006466257385909557, + 0.009896170347929, + 0.00929190032184124, + 0.006802923046052456, + 0.016971668228507042, + 0.014422126114368439, + 0.011740593239665031, + 0.013829163275659084, + 0.013377870433032513, + 0.011139556765556335, + 0.007061503827571869, + 0.0059995162300765514, + 0.006447601597756147, + 0.0824027806520462, + 0.012841511517763138, + 0.07957526296377182, + 0.006706038024276495, + 0.009821448475122452, + 0.007013247348368168, + 0.01036952342838049, + 0.008181205950677395, + 0.03495831787586212, + 0.010307145304977894, + 0.03427327051758766, + 0.0033644901122897863, + 0.002548233140259981, + 0.0025126831606030464, + 0.002752465195953846, + 0.003819245146587491, + 0.004055823665112257, + 0.00595890125259757, + 0.0035631253849714994, + 0.004931130446493626, + 0.004266475792974234, + 0.005081298295408487, + 0.002974849194288254, + 0.007537933997809887, + 0.005078359507024288, + 0.004528367426246405, + 0.005391384474933147, + 0.011330883018672466, + 0.014842086471617222, + 0.01217059139162302, + 0.00818805955350399, + 0.023122867569327354, + 0.014037045650184155, + 0.007345388177782297, + 0.012687663547694683, + 0.01553197018802166, + 0.0063319033943116665, + 0.007925036363303661, + 0.009054398164153099, + 0.0071245888248085976, + 0.009329486638307571, + 0.005840296391397715, + 0.00755193829536438, + 0.021992791444063187, + 0.006693718954920769, + 0.013296082615852356, + 0.010180572047829628, + 0.004566722549498081, + 0.010934436693787575, + 0.00625898502767086, + 0.010742386803030968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.07004838436841965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052674319595098495 + ], + "opposite_logprob_tokens": [ + -5.25526762008667 + ], + "logprob_reward": -0.0052674319595098495, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052674319595098495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13838452100753784, + 0.1452065408229828, + 0.0879327580332756, + 0.2071075439453125, + 0.16764996945858002, + 0.04685816541314125, + 0.05811483785510063, + 0.05446343123912811, + 0.07274125516414642, + 0.0215410478413105 + ], + "credit_map": [ + 0.009283522143959999, + 0.018910130485892296, + 0.020764701068401337, + 0.0100640207529068, + 0.017169324681162834, + 0.02402826026082039, + 0.01405567117035389, + 0.024108892306685448, + 0.007026376202702522, + 0.010964944027364254, + 0.01604829542338848, + 0.013521462678909302, + 0.04150077700614929, + 0.022320136427879333, + 0.018557710573077202, + 0.015266837552189827, + 0.006590177770704031, + 0.011369943618774414, + 0.00573444738984108, + 0.006731976754963398, + 0.01859096810221672, + 0.013996638357639313, + 0.01165519468486309, + 0.013263413682579994, + 0.011839864775538445, + 0.011831214651465416, + 0.01049642451107502, + 0.011605892330408096, + 0.01110587827861309, + 0.06559016555547714, + 0.018796665593981743, + 0.06584141403436661, + 0.011680901050567627, + 0.014370753429830074, + 0.011293131858110428, + 0.02878977544605732, + 0.016348419710993767, + 0.03171859681606293, + 0.020997678861021996, + 0.03245071321725845, + 0.005527370143681765, + 0.0042088646441698074, + 0.002718111500144005, + 0.004211581777781248, + 0.006220031529664993, + 0.007658595219254494, + 0.008940307423472404, + 0.007373302709311247, + 0.008399350568652153, + 0.00455879420042038, + 0.008065720088779926, + 0.004889717325568199, + 0.005217482335865498, + 0.008529573678970337, + 0.006845314987003803, + 0.011608880013227463, + 0.010273061692714691, + 0.0052172779105603695, + 0.005526335909962654, + 0.003717772662639618, + 0.005779188591986895, + 0.009618379175662994, + 0.003985004965215921, + 0.010346408933401108, + 0.00983394868671894, + 0.007934030145406723, + 0.008378379046916962, + 0.011812916956841946, + 0.008693188428878784, + 0.010354309342801571, + 0.007752010598778725, + 0.007982468232512474, + 0.004900046158581972, + 0.0014388406416401267, + 0.0028309114277362823, + 0.003057257505133748, + 0.001525946194306016, + 0.0024544440675526857, + 0.0022615683265030384, + 0.0030720329377800226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.0052674319595098495 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25207653641700745 + ], + "opposite_logprob_tokens": [ + -1.502076506614685 + ], + "logprob_reward": -0.25207653641700745, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207653641700745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13845032453536987, + 0.07479256391525269, + 0.0801621749997139, + 0.17042392492294312, + 0.07603263854980469, + 0.17186108231544495, + 0.18926973640918732, + 0.03887667879462242, + 0.0414072684943676, + 0.018723594024777412 + ], + "credit_map": [ + 0.011984857730567455, + 0.02016501873731613, + 0.01900097168982029, + 0.013175304047763348, + 0.01781231351196766, + 0.017782503738999367, + 0.0191279798746109, + 0.019401393830776215, + 0.0041384086944162846, + 0.004579216707497835, + 0.0066760387271642685, + 0.00433021504431963, + 0.007722172886133194, + 0.020367078483104706, + 0.007826893590390682, + 0.019152548164129257, + 0.00540159922093153, + 0.00669964961707592, + 0.005780074745416641, + 0.006761002354323864, + 0.006198476999998093, + 0.02118871547281742, + 0.00794164463877678, + 0.020191019400954247, + 0.011019323021173477, + 0.007424122653901577, + 0.007295557297766209, + 0.004556356929242611, + 0.004229012876749039, + 0.06532516330480576, + 0.006954725831747055, + 0.06361966580152512, + 0.00845826044678688, + 0.007419987116008997, + 0.004879197105765343, + 0.011258364655077457, + 0.006397789344191551, + 0.012686575762927532, + 0.011215698905289173, + 0.013716774061322212, + 0.026184387505054474, + 0.011342533864080906, + 0.018899422138929367, + 0.022061627358198166, + 0.021993964910507202, + 0.025040078908205032, + 0.0262370053678751, + 0.02010207809507847, + 0.02908160910010338, + 0.015566540881991386, + 0.025674620643258095, + 0.013408860191702843, + 0.023038242012262344, + 0.029063057154417038, + 0.023033523932099342, + 0.03040330298244953, + 0.006012343801558018, + 0.0028241437394171953, + 0.005377808585762978, + 0.0027247497346252203, + 0.0028677231166511774, + 0.006577603984624147, + 0.002807219745591283, + 0.009685088880360126, + 0.0054100146517157555, + 0.00455732736736536, + 0.0063932365737855434, + 0.005432958714663982, + 0.005417351145297289, + 0.005915146786719561, + 0.00517444359138608, + 0.003106796182692051, + 0.0034279413521289825, + 0.0015881212893873453, + 0.0022259692195802927, + 0.0021112519316375256, + 0.00182257906999439, + 0.0035314818378537893, + 0.0011372732697054744, + 0.0028789793141186237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.25207653641700745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022019920870661736 + ], + "opposite_logprob_tokens": [ + -6.127202033996582 + ], + "logprob_reward": -0.0022019920870661736, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022019920870661736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07563885301351547, + 0.08697686344385147, + 0.04648962989449501, + 0.25569334626197815, + 0.202158123254776, + 0.04578506946563721, + 0.0918828472495079, + 0.10176750272512436, + 0.06397412717342377, + 0.029633691534399986 + ], + "credit_map": [ + 0.005847118329256773, + 0.0070114824920892715, + 0.013366411440074444, + 0.008425862528383732, + 0.009889116510748863, + 0.011223428882658482, + 0.009057809598743916, + 0.01081761159002781, + 0.00707469554618001, + 0.00737902894616127, + 0.01049343217164278, + 0.006390469614416361, + 0.0061122216284275055, + 0.021959859877824783, + 0.008183974772691727, + 0.01938316598534584, + 0.003084002062678337, + 0.004484920762479305, + 0.005910009145736694, + 0.004717255011200905, + 0.009202934801578522, + 0.005844751838594675, + 0.008125529624521732, + 0.005120226182043552, + 0.015303056687116623, + 0.015920624136924744, + 0.018249783664941788, + 0.019287677481770515, + 0.01505831629037857, + 0.07937188446521759, + 0.015295232646167278, + 0.07720675319433212, + 0.014190067537128925, + 0.011799351312220097, + 0.008003885857760906, + 0.01382366195321083, + 0.013243124820291996, + 0.06080741435289383, + 0.018739823251962662, + 0.06155078858137131, + 0.009946197271347046, + 0.0026104359421879053, + 0.003301978809759021, + 0.004289410542696714, + 0.0032919063232839108, + 0.007533051539212465, + 0.0071228607557713985, + 0.00768922409042716, + 0.012836596928536892, + 0.01060605887323618, + 0.008506816811859608, + 0.007452350575476885, + 0.016907496377825737, + 0.012810606509447098, + 0.010771889239549637, + 0.011991024948656559, + 0.012370768934488297, + 0.00852380320429802, + 0.01106787659227848, + 0.00650749821215868, + 0.021206436678767204, + 0.01611356995999813, + 0.007566856686025858, + 0.018410680815577507, + 0.015960993245244026, + 0.0069467839784920216, + 0.007154664024710655, + 0.006680967751890421, + 0.00650151027366519, + 0.009273504838347435, + 0.00519675062969327, + 0.006258948240429163, + 0.005847017280757427, + 0.0018223304068669677, + 0.0037698044907301664, + 0.004374849610030651, + 0.0022656628862023354, + 0.004139854572713375, + 0.002946777269244194, + 0.00446739187464118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.0022019920870661736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-contrast digital look, the slow motion of a 360-degree orbiting camera captures a moment of intense repulsion, frozen in extreme slow motion within a desolate black void. At the center of the frame, a heavy, solid block of black ink and a delicate white porcelain pen holder are forcefully flung away from each other after an unseen, powerful collision. The inkstone, a rectangular block composed of matte, dark slate, slowly rotates along its axis, its dense shape remaining largely intact, save for a small, newly chiseled corner that gleams in the harsh light. In contrast, the cylindrical porcelain pen holder shatters catastrophically, its thin, lustrous white fragments ejected outward in an expanding sphere. A beautiful, chaotic cloud of fine white ceramic dust erupts from the point of impact, catching the single, powerful keylight and emitting an ethereal glow, while sharper, grainy black particles from the ink shoot outward from this cloud in more direct trajectories, like microscopic obsidian fragments. As the camera gracefully orbits this levitating explosion, high-contrast lighting sculpts out every stunning detail—the frantic spiderweb textures on shattered porcelain, the smooth, cool surface of ink, and individual specks of dust—creating a scene of controlled chaos of breathtaking beauty and technical precision.", + "seed": 8029, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_14/index_8761_seed_8029_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.18526925100013614, + "vlm_logprob": -0.18526925100013614, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.18526925100013614, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006776447407901287 + ], + "opposite_logprob_tokens": [ + -5.256776332855225 + ], + "logprob_reward": -0.006776447407901287, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.006776447407901287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14530207216739655, + 0.19251370429992676, + 0.047227419912815094, + 0.06550576537847519, + 0.10968214273452759, + 0.14359606802463531, + 0.050249651074409485, + 0.08037228137254715, + 0.0516362227499485, + 0.11391468346118927 + ], + "credit_map": [ + 0.017324231564998627, + 0.01136284414678812, + 0.021939266473054886, + 0.016831811517477036, + 0.01550924964249134, + 0.023807765915989876, + 0.018560193479061127, + 0.01996670663356781, + 0.021537665277719498, + 0.007727855816483498, + 0.05846632272005081, + 0.008010934107005596, + 0.009760689921677113, + 0.05796261504292488, + 0.013177449814975262, + 0.015870170667767525, + 0.004670547321438789, + 0.004192635416984558, + 0.0073781064711511135, + 0.0026551096234470606, + 0.004576937761157751, + 0.009364559315145016, + 0.005877351853996515, + 0.008512173779308796, + 0.007646202575415373, + 0.009895761497318745, + 0.009621787816286087, + 0.007517529651522636, + 0.006486264523118734, + 0.008329074829816818, + 0.0063545615412294865, + 0.009654583409428596, + 0.01281492505222559, + 0.011675508692860603, + 0.014946261420845985, + 0.019478190690279007, + 0.007814625278115273, + 0.015592485666275024, + 0.011676878668367863, + 0.015683263540267944, + 0.03152846917510033, + 0.019381029531359673, + 0.021856704726815224, + 0.011957474984228611, + 0.008891411125659943, + 0.01743357628583908, + 0.014571281149983406, + 0.017976121976971626, + 0.0085755018517375, + 0.007370868232101202, + 0.011327731423079967, + 0.004523066338151693, + 0.0055722338147461414, + 0.004861458670347929, + 0.004918245133012533, + 0.00310054374858737, + 0.012713386677205563, + 0.012792273424565792, + 0.016277605667710304, + 0.009366964921355247, + 0.007494844030588865, + 0.010956966318190098, + 0.005625095218420029, + 0.005145142786204815, + 0.010872662998735905, + 0.006451859604567289, + 0.01430621650069952, + 0.006532703060656786, + 0.002769355196505785, + 0.004259994253516197, + 0.003708784468472004, + 0.0027346457354724407, + 0.03305286541581154, + 0.03160947933793068, + 0.011304656974971294, + 0.009389939717948437, + 0.006973253097385168, + 0.009479722939431667, + 0.006041395012289286, + 0.006063369102776051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.006776447407901287 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153866708278656 + ], + "opposite_logprob_tokens": [ + -1.7015386819839478 + ], + "logprob_reward": -0.20153866708278656, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153866708278656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22728762030601501, + 0.16231811046600342, + 0.05834430083632469, + 0.039175864309072495, + 0.2917812466621399, + 0.07449911534786224, + 0.030818089842796326, + 0.025851910933852196, + 0.01711564138531685, + 0.07280809432268143 + ], + "credit_map": [ + 0.019141005352139473, + 0.02483273483812809, + 0.04713829606771469, + 0.02121923118829727, + 0.02112511172890663, + 0.04679085314273834, + 0.01999298296868801, + 0.02704739384353161, + 0.008507280610501766, + 0.005305553320795298, + 0.056920379400253296, + 0.0066238874569535255, + 0.010144499130547047, + 0.05720372870564461, + 0.010787424631416798, + 0.006825343240052462, + 0.005996326915919781, + 0.005278990138322115, + 0.010098213329911232, + 0.004902814514935017, + 0.0051255193538963795, + 0.01011558435857296, + 0.006997344084084034, + 0.009829509072005749, + 0.004673892632126808, + 0.006166566628962755, + 0.005671776831150055, + 0.0033609673846513033, + 0.003110147314146161, + 0.005696607753634453, + 0.0031992655713111162, + 0.007296644151210785, + 0.04031546041369438, + 0.0367082841694355, + 0.0419343002140522, + 0.05771815404295921, + 0.021210335195064545, + 0.03346409276127815, + 0.03384153172373772, + 0.026589082553982735, + 0.015306195244193077, + 0.00818447396159172, + 0.010113481432199478, + 0.010232610628008842, + 0.006960142403841019, + 0.009292439557611942, + 0.007220996078103781, + 0.007188775110989809, + 0.00491989916190505, + 0.0053350478410720825, + 0.0061884233728051186, + 0.0031241707038134336, + 0.002958307508379221, + 0.003911541774868965, + 0.0023415738251060247, + 0.0020391270518302917, + 0.004987584892660379, + 0.004728422500193119, + 0.005674590356647968, + 0.003063699696213007, + 0.001955542014911771, + 0.001982899149879813, + 0.0021938385907560587, + 0.0012653334997594357, + 0.0032603780273348093, + 0.001889494713395834, + 0.0045268600806593895, + 0.002417868934571743, + 0.0015113475965335965, + 0.0013033944414928555, + 0.0013424514327198267, + 0.0008638477302156389, + 0.024326639249920845, + 0.02232637256383896, + 0.007439760025590658, + 0.00447517354041338, + 0.004084975924342871, + 0.005302270874381065, + 0.003015841357409954, + 0.0018370524048805237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.20153866708278656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004107373300939798 + ], + "opposite_logprob_tokens": [ + -5.504107475280762 + ], + "logprob_reward": -0.004107373300939798, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004107373300939798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12594068050384521, + 0.29700449109077454, + 0.08817818760871887, + 0.11767648160457611, + 0.060176458209753036, + 0.052460163831710815, + 0.04174937680363655, + 0.0549604557454586, + 0.052369408309459686, + 0.10948435962200165 + ], + "credit_map": [ + 0.013260304927825928, + 0.015116170048713684, + 0.019826464354991913, + 0.013318795710802078, + 0.010328147560358047, + 0.018244894221425056, + 0.013746526092290878, + 0.022099370136857033, + 0.015978196635842323, + 0.011039105243980885, + 0.10242260247468948, + 0.012771041132509708, + 0.012931873090565205, + 0.10461415350437164, + 0.02123156748712063, + 0.016015948727726936, + 0.009028490632772446, + 0.006243361625820398, + 0.013571478426456451, + 0.0072053880430758, + 0.006545829586684704, + 0.01229737140238285, + 0.012032361701130867, + 0.021253904327750206, + 0.014202439226210117, + 0.019131962209939957, + 0.017766421660780907, + 0.008291380479931831, + 0.013129042461514473, + 0.01632649637758732, + 0.009697840549051762, + 0.01913088746368885, + 0.006344223860651255, + 0.007271862123161554, + 0.010801447555422783, + 0.004170581698417664, + 0.005859241355210543, + 0.009738704189658165, + 0.0065109278075397015, + 0.009479464963078499, + 0.009850548580288887, + 0.006691498216241598, + 0.00808100588619709, + 0.007002669386565685, + 0.004865961614996195, + 0.006462080404162407, + 0.004651237744837999, + 0.004855161067098379, + 0.0073939659632742405, + 0.006874025333672762, + 0.0067557343281805515, + 0.006496984977275133, + 0.0035779287572950125, + 0.004132581874728203, + 0.004106990061700344, + 0.0024111648090183735, + 0.010471571236848831, + 0.009207055903971195, + 0.013566435314714909, + 0.005542728118598461, + 0.004939590580761433, + 0.004542665556073189, + 0.004343575332313776, + 0.002346829744055867, + 0.009279092773795128, + 0.007602126803249121, + 0.011019767262041569, + 0.007523273583501577, + 0.004096530843526125, + 0.005503013730049133, + 0.003929666243493557, + 0.003415933810174465, + 0.02486441470682621, + 0.027014166116714478, + 0.011970619671046734, + 0.008921596221625805, + 0.008709828369319439, + 0.012435493990778923, + 0.0074464683420956135, + 0.008121767081320286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.004107373300939798 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933398246765137 + ], + "opposite_logprob_tokens": [ + -0.6933398246765137 + ], + "logprob_reward": -0.6933398246765137, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933398246765137, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14673548936843872, + 0.10203798115253448, + 0.049572497606277466, + 0.04684446379542351, + 0.04126780480146408, + 0.09900090098381042, + 0.19906388223171234, + 0.15795424580574036, + 0.03062472864985466, + 0.12689803540706635 + ], + "credit_map": [ + 0.013510668650269508, + 0.017032913863658905, + 0.02119811438024044, + 0.018737738952040672, + 0.021634921431541443, + 0.012003881856799126, + 0.03100057691335678, + 0.011616670526564121, + 0.0056887357495725155, + 0.004233935382217169, + 0.03100421465933323, + 0.007504486944526434, + 0.007287336513400078, + 0.03303387016057968, + 0.007340834941715002, + 0.0059445695951581, + 0.004967575892806053, + 0.004500850103795528, + 0.007536202669143677, + 0.003467998467385769, + 0.002879880368709564, + 0.007924213074147701, + 0.005673407576978207, + 0.012622371315956116, + 0.005747104063630104, + 0.0062552099116146564, + 0.00866784155368805, + 0.002973903901875019, + 0.004947783425450325, + 0.0069587137550115585, + 0.0045447275042533875, + 0.006749178748577833, + 0.005088373553007841, + 0.003897029208019376, + 0.007037458010017872, + 0.0025935936719179153, + 0.004895305726677179, + 0.005820257589221001, + 0.0038385549560189247, + 0.008097232319414616, + 0.02838948927819729, + 0.010388643480837345, + 0.01699610985815525, + 0.008656878024339676, + 0.006214683875441551, + 0.013070636428892612, + 0.00711798295378685, + 0.008166477084159851, + 0.03152170404791832, + 0.04115738719701767, + 0.029907982796430588, + 0.030527405440807343, + 0.02053656242787838, + 0.01671474613249302, + 0.016871225088834763, + 0.011826866306364536, + 0.02787245623767376, + 0.025312010198831558, + 0.0372089147567749, + 0.018636269494891167, + 0.013624513521790504, + 0.018956560641527176, + 0.008667060174047947, + 0.007676462177187204, + 0.005071721505373716, + 0.004227651748806238, + 0.005707939621061087, + 0.004897926934063435, + 0.0023818330373615026, + 0.002726502949371934, + 0.003162371227517724, + 0.002448780694976449, + 0.03466193005442619, + 0.03665066137909889, + 0.013279964216053486, + 0.00978437066078186, + 0.00965873058885336, + 0.00992539431899786, + 0.008325318805873394, + 0.004611670505255461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.6933398246765137 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, solid rectangular block of matte dark slate (inkstone) visible at the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cylindrical white porcelain pen holder visible that shatters into thin, lustrous white fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow 360-degree orbit around the central collision point?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the objects and debris captured in extreme slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583942532539368 + ], + "opposite_logprob_tokens": [ + -3.8955838680267334 + ], + "logprob_reward": -0.020583942532539368, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583942532539368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08619926124811172, + 0.13533331453800201, + 0.05985792353749275, + 0.06952822208404541, + 0.07376448810100555, + 0.07526036351919174, + 0.17878630757331848, + 0.2094089686870575, + 0.07520375400781631, + 0.03665737807750702 + ], + "credit_map": [ + 0.010730814188718796, + 0.008213212713599205, + 0.014051037840545177, + 0.008081122301518917, + 0.009790943935513496, + 0.011183034628629684, + 0.012403908185660839, + 0.01174518745392561, + 0.007741378154605627, + 0.006951652001589537, + 0.0405270978808403, + 0.008276162669062614, + 0.006923051550984383, + 0.04512564837932587, + 0.010002927854657173, + 0.009785391390323639, + 0.0060727777890861034, + 0.004168279934674501, + 0.009394564665853977, + 0.005266600288450718, + 0.004850140307098627, + 0.009677967056632042, + 0.012006786651909351, + 0.008420808240771294, + 0.008001331239938736, + 0.010211579501628876, + 0.009462870657444, + 0.005059791263192892, + 0.007063000462949276, + 0.007215012330561876, + 0.011582744307816029, + 0.010931896977126598, + 0.011750095523893833, + 0.009069631807506084, + 0.011935633607208729, + 0.010641206055879593, + 0.005125416908413172, + 0.00702148349955678, + 0.00962417759001255, + 0.008596837520599365, + 0.014464977197349072, + 0.009692555293440819, + 0.009810922667384148, + 0.007359674200415611, + 0.006861918140202761, + 0.010052566416561604, + 0.007278796751052141, + 0.009738956578075886, + 0.03186112269759178, + 0.030566349625587463, + 0.03428175672888756, + 0.024403534829616547, + 0.016387013718485832, + 0.01748581975698471, + 0.012663014233112335, + 0.011137695051729679, + 0.032741568982601166, + 0.03171675279736519, + 0.03944329544901848, + 0.023098528385162354, + 0.019416101276874542, + 0.03287191316485405, + 0.015317137353122234, + 0.01480367686599493, + 0.013908027671277523, + 0.010246437974274158, + 0.017621096223592758, + 0.00818806141614914, + 0.0049522132612764835, + 0.010199303738772869, + 0.005572649650275707, + 0.004515965469181538, + 0.006126427091658115, + 0.007000488229095936, + 0.004726411774754524, + 0.006148266140371561, + 0.0030291404109448195, + 0.0029408251866698265, + 0.0038280044682323933, + 0.002857811516150832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.020583942532539368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_2/index_tensor([8280, 8280])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_2/index_tensor([8280, 8280])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f33472360ca4aec51972e41bbef8cbd0e8182771 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_2/index_tensor([8280, 8280])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, long shot, slowly pulling back, captures a breathtaking scene inside an Olympic-standard indoor swimming pool. At the center of the frame, a muscular male swimmer, clad in a white swim cap, black goggles, and tight-fitting black racing trunks, surges through the clear, azure water with an almost ferocious force. Spotlights from the gymnasium's ceiling cast bright, sharp beams of light across the surface, creating a constantly shifting, distorted play of light and shadow beneath the turbulent water. Each of his arm strokes slices through the water with incredible force, seemingly ripping through the very surface, sending up massive white splashes and turbulent currents. Under the surface, countless tiny bubbles, stirred up by his powerful leg kicks, form tumbling white trails, creating the visual effect of a violent submarine landslide. The immense force of this impact causes the entire pool's surface to vibrate violently, waves radiating from him crashing against the edge, and the clear black lane lines distort and sway beneath the surface. The entire image is filled with the tension of power, speed and extreme concentration. The contrast between the deep blue and the dazzling white foam perfectly captures the breathtaking epic feeling and tense atmosphere of the Olympic final.", + "seed": 25367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_2/index_8280_seed_25367_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12298198696225882, + "vlm_logprob": -0.12298198696225882, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.12298198696225882, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13923892378807068 + ], + "opposite_logprob_tokens": [ + -2.6392388343811035 + ], + "logprob_reward": -0.13923892378807068, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.13923892378807068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47024983167648315, + 0.1812562644481659, + 0.07890476286411285, + 0.06367361545562744, + 0.0282671470195055, + 0.05642280355095863, + 0.057080723345279694, + 0.01327322144061327, + 0.027478624135255814, + 0.023392997682094574 + ], + "credit_map": [ + 0.10437038540840149, + 0.04516033083200455, + 0.03822650760412216, + 0.038093652576208115, + 0.047015879303216934, + 0.03686677664518356, + 0.051893457770347595, + 0.10862283408641815, + 0.03695178031921387, + 0.029242081567645073, + 0.018524404615163803, + 0.019312188029289246, + 0.016749078407883644, + 0.011666906997561455, + 0.023233234882354736, + 0.02557658776640892, + 0.0063250502571463585, + 0.02037716470658779, + 0.005953384563326836, + 0.005438880994915962, + 0.008610379882156849, + 0.0072951060719788074, + 0.006816353648900986, + 0.018088448792696, + 0.0070208655670285225, + 0.007720476016402245, + 0.007734566926956177, + 0.010215409100055695, + 0.00747649185359478, + 0.005298107862472534, + 0.010365787893533707, + 0.007841916754841805, + 0.00252630771137774, + 0.0030185626819729805, + 0.0037257468793541193, + 0.0033886716701090336, + 0.004833269864320755, + 0.0045588938519358635, + 0.002999374642968178, + 0.0032163208816200495, + 0.006990300957113504, + 0.005233765114098787, + 0.0070090158842504025, + 0.006035084370523691, + 0.006617567036300898, + 0.01171173620969057, + 0.005892596207559109, + 0.006932741031050682, + 0.011495430022478104, + 0.004243294708430767, + 0.009303783997893333, + 0.01010224036872387, + 0.00586438924074173, + 0.0036305247340351343, + 0.006512482650578022, + 0.00592857925221324, + 0.0013165578711777925, + 0.001545047271065414, + 0.0020771201234310865, + 0.003171436255797744, + 0.0012905355542898178, + 0.001134665566496551, + 0.0016388639342039824, + 0.001098995329812169, + 0.006937652360647917, + 0.0014544568257406354, + 0.0075465780682861805, + 0.0020451426971703768, + 0.00173345603980124, + 0.00261736405082047, + 0.002062703249976039, + 0.0030812721233814955, + 0.0026132219936698675, + 0.001724181929603219, + 0.004357853904366493, + 0.005280300509184599, + 0.0018950891681015491, + 0.0021767430007457733, + 0.0025949289556592703, + 0.002750677289441228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.13923892378807068 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870583474636078 + ], + "opposite_logprob_tokens": [ + -1.1370583772659302 + ], + "logprob_reward": -0.3870583474636078, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870583474636078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19958584010601044, + 0.26592257618904114, + 0.14144620299339294, + 0.12461073696613312, + 0.05714255943894386, + 0.0399068184196949, + 0.05284387990832329, + 0.08516097068786621, + 0.00831504724919796, + 0.025065410882234573 + ], + "credit_map": [ + 0.05011066421866417, + 0.01846516691148281, + 0.01453658752143383, + 0.014367751777172089, + 0.014181076548993587, + 0.013829754665493965, + 0.020620528608560562, + 0.0534742996096611, + 0.03148733451962471, + 0.05638255178928375, + 0.012536470778286457, + 0.017618417739868164, + 0.017319295555353165, + 0.08416730910539627, + 0.018230009824037552, + 0.02818121388554573, + 0.014512965455651283, + 0.03534597530961037, + 0.008655291981995106, + 0.008877582848072052, + 0.01630370132625103, + 0.011167516931891441, + 0.009272345341742039, + 0.03731081262230873, + 0.007900325581431389, + 0.012247106060385704, + 0.014411022886633873, + 0.0232671070843935, + 0.014024142175912857, + 0.01175280474126339, + 0.02532135508954525, + 0.015686865895986557, + 0.005768268369138241, + 0.00513015640899539, + 0.005638669244945049, + 0.00424245884642005, + 0.014047118835151196, + 0.009553714655339718, + 0.006844366434961557, + 0.005917803384363651, + 0.004759644623845816, + 0.002583072753623128, + 0.004439228679984808, + 0.0035645051393657923, + 0.004473645240068436, + 0.011093216016888618, + 0.0028489436954259872, + 0.00614456320181489, + 0.005317730363458395, + 0.003705883165821433, + 0.011800184845924377, + 0.013419041410088539, + 0.003171498654410243, + 0.0042809913866221905, + 0.0055029247887432575, + 0.005645626690238714, + 0.014538202434778214, + 0.005756651051342487, + 0.010200499556958675, + 0.03150397539138794, + 0.0031558640766888857, + 0.0037670438177883625, + 0.010271070525050163, + 0.005967666395008564, + 0.0008837510249577463, + 0.0007386409561149776, + 0.0009835708187893033, + 0.0008736195741221309, + 0.0012558571761474013, + 0.0008430310990661383, + 0.001657153363339603, + 0.0010794230038300157, + 0.0017887346912175417, + 0.002500408561900258, + 0.003694533836096525, + 0.0037747754249721766, + 0.002450014464557171, + 0.002262875670567155, + 0.006709012668579817, + 0.001885056379251182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.3870583474636078 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207091733813286 + ], + "opposite_logprob_tokens": [ + -2.812070846557617 + ], + "logprob_reward": -0.06207091733813286, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207091733813286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.287109911441803, + 0.1266830712556839, + 0.06104446202516556, + 0.08283006399869919, + 0.04657283425331116, + 0.08563150465488434, + 0.15325021743774414, + 0.032905422151088715, + 0.10218927264213562, + 0.021783247590065002 + ], + "credit_map": [ + 0.06309283524751663, + 0.022843975573778152, + 0.02706102281808853, + 0.020299691706895828, + 0.02034890092909336, + 0.025724666193127632, + 0.03018934279680252, + 0.07754945755004883, + 0.023122472688555717, + 0.01473036129027605, + 0.0065705059096217155, + 0.008321805857121944, + 0.010934998281300068, + 0.04004571959376335, + 0.00938368495553732, + 0.013573517091572285, + 0.006526837591081858, + 0.007235878147184849, + 0.009104598313570023, + 0.006852719932794571, + 0.00947059690952301, + 0.0052604712545871735, + 0.009711815044283867, + 0.006881538312882185, + 0.008653792552649975, + 0.01104817260056734, + 0.009644785895943642, + 0.011041685938835144, + 0.014310858212411404, + 0.00923929363489151, + 0.010242244228720665, + 0.008649221621453762, + 0.004861123859882355, + 0.0030446676537394524, + 0.00553172780200839, + 0.0054001878015697, + 0.010068872012197971, + 0.007134478539228439, + 0.005719279870390892, + 0.004812495317310095, + 0.009596667252480984, + 0.007466546725481749, + 0.011638496071100235, + 0.008840464055538177, + 0.008835279382765293, + 0.013663857243955135, + 0.007909530773758888, + 0.017680654302239418, + 0.03228922560811043, + 0.013472716324031353, + 0.02114083804190159, + 0.02374446950852871, + 0.016705336049199104, + 0.011316505260765553, + 0.018325157463550568, + 0.016255952417850494, + 0.006496497895568609, + 0.0025815358385443687, + 0.003653810592368245, + 0.010389014147222042, + 0.0014661183813586831, + 0.0014727514935657382, + 0.004586929455399513, + 0.0022587610874325037, + 0.027901260182261467, + 0.00513388030230999, + 0.027625273913145065, + 0.008337387815117836, + 0.006070530973374844, + 0.007302969228476286, + 0.011901883408427238, + 0.007916085422039032, + 0.0025530315469950438, + 0.002687396015971899, + 0.0027345065027475357, + 0.003584540681913495, + 0.0018025895114988089, + 0.0027416562661528587, + 0.0031058243475854397, + 0.0025736994575709105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.06207091733813286 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059316821396351 + ], + "opposite_logprob_tokens": [ + -3.8955931663513184 + ], + "logprob_reward": -0.02059316821396351, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059316821396351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18463806807994843, + 0.14762552082538605, + 0.10352403670549393, + 0.0703037679195404, + 0.052260782569646835, + 0.029061267152428627, + 0.10285484790802002, + 0.16481828689575195, + 0.028063645586371422, + 0.11684972792863846 + ], + "credit_map": [ + 0.020547794178128242, + 0.03079911135137081, + 0.02338547818362713, + 0.016708312556147575, + 0.027589375153183937, + 0.016921456903219223, + 0.018843162804841995, + 0.029843388125300407, + 0.026172099635004997, + 0.03183173015713692, + 0.006128148175776005, + 0.007045278325676918, + 0.00868857279419899, + 0.05136967450380325, + 0.006569843273609877, + 0.009820175357162952, + 0.009213297627866268, + 0.027487359941005707, + 0.0069201188161969185, + 0.005308746360242367, + 0.010226418264210224, + 0.007176180835813284, + 0.00731640262529254, + 0.02987552620470524, + 0.008028780110180378, + 0.011205288581550121, + 0.0083039291203022, + 0.010133184492588043, + 0.00990758091211319, + 0.005835323594510555, + 0.007972855120897293, + 0.008916831575334072, + 0.005922222044318914, + 0.005188569892197847, + 0.006568367127329111, + 0.004421916324645281, + 0.010915614664554596, + 0.006821902003139257, + 0.006144868675619364, + 0.006277323700487614, + 0.0031744297593832016, + 0.0023070781026035547, + 0.003799523925408721, + 0.0030900980345904827, + 0.0031278603710234165, + 0.006284398026764393, + 0.0033620167523622513, + 0.003915863577276468, + 0.01599285937845707, + 0.008835048414766788, + 0.014594663865864277, + 0.016245270147919655, + 0.00977769959717989, + 0.01205387618392706, + 0.014038587920367718, + 0.011316850781440735, + 0.029390381649136543, + 0.011323046870529652, + 0.01883983425796032, + 0.061797671020030975, + 0.006224562413990498, + 0.00571596110239625, + 0.02042207680642605, + 0.011104763485491276, + 0.006463943514972925, + 0.002146609593182802, + 0.005796887446194887, + 0.0032669345382601023, + 0.0030096801929175854, + 0.0025335834361612797, + 0.00257753930054605, + 0.002268468029797077, + 0.011477087624371052, + 0.012371341697871685, + 0.01272030919790268, + 0.0152917904779315, + 0.014893566258251667, + 0.009050494991242886, + 0.03141907975077629, + 0.009626069106161594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.02059316821396351 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948578007519245 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948578007519245, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948578007519245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41755375266075134, + 0.08075281232595444, + 0.15028312802314758, + 0.12337274104356766, + 0.014115612022578716, + 0.03638403117656708, + 0.037231553345918655, + 0.029561379924416542, + 0.07571301609277725, + 0.035031966865062714 + ], + "credit_map": [ + 0.07838407903909683, + 0.0482122041285038, + 0.04350917041301727, + 0.05231526121497154, + 0.03470500931143761, + 0.03395681828260422, + 0.04021349176764488, + 0.08625772595405579, + 0.012857625260949135, + 0.00873265415430069, + 0.008220040239393711, + 0.008403236977756023, + 0.007574081886559725, + 0.008532222360372543, + 0.015824802219867706, + 0.010608154349029064, + 0.017871668562293053, + 0.026886457577347755, + 0.01122249010950327, + 0.014213747344911098, + 0.019551564007997513, + 0.013252378441393375, + 0.0203287061303854, + 0.02695612423121929, + 0.01126584131270647, + 0.01773390732705593, + 0.014633317478001118, + 0.01573125645518303, + 0.018344096839427948, + 0.012742100283503532, + 0.020860252901911736, + 0.012061967514455318, + 0.0017673231195658445, + 0.0013994092587381601, + 0.001784993102774024, + 0.001872918102890253, + 0.0023163803853094578, + 0.0017063815612345934, + 0.0014812789158895612, + 0.0017869272269308567, + 0.0033297010231763124, + 0.0031024825293570757, + 0.0035163587890565395, + 0.0028727995231747627, + 0.0037858239375054836, + 0.009272073395550251, + 0.0029026020783931017, + 0.007602190598845482, + 0.005643304903060198, + 0.002489053877070546, + 0.007307976018637419, + 0.007722028531134129, + 0.0031784202437847853, + 0.002748222788795829, + 0.0043925573118031025, + 0.003749990602955222, + 0.004760729148983955, + 0.002802935428917408, + 0.0034677430521696806, + 0.010367431677877903, + 0.0011763679794967175, + 0.0010488084517419338, + 0.003942141309380531, + 0.0019952235743403435, + 0.022233864292502403, + 0.0036989992950111628, + 0.022749213501811028, + 0.005345451179891825, + 0.004080828744918108, + 0.0042127701453864574, + 0.0072228261269629, + 0.006169059779495001, + 0.004352851305156946, + 0.006326143629848957, + 0.004368084482848644, + 0.004261295311152935, + 0.004317425657063723, + 0.0026476436760276556, + 0.005003074184060097, + 0.0037554502487182617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.005948578007519245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long shot, slowly pulling back, captures a breathtaking scene inside an Olympic-standard indoor swimming pool. At the center of the frame, a muscular male swimmer, clad in a white swim cap, black goggles, and tight-fitting black racing trunks, surges through the clear, azure water with an almost ferocious force. Spotlights from the gymnasium's ceiling cast bright, sharp beams of light across the surface, creating a constantly shifting, distorted play of light and shadow beneath the turbulent water. Each of his arm strokes slices through the water with incredible force, seemingly ripping through the very surface, sending up massive white splashes and turbulent currents. Under the surface, countless tiny bubbles, stirred up by his powerful leg kicks, form tumbling white trails, creating the visual effect of a violent submarine landslide. The immense force of this impact causes the entire pool's surface to vibrate violently, waves radiating from him crashing against the edge, and the clear black lane lines distort and sway beneath the surface. The entire image is filled with the tension of power, speed and extreme concentration. The contrast between the deep blue and the dazzling white foam perfectly captures the breathtaking epic feeling and tense atmosphere of the Olympic final.", + "seed": 25368, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_2/index_8280_seed_25368_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1002663753926754, + "vlm_logprob": -0.1002663753926754, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.1002663753926754, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10260503739118576 + ], + "opposite_logprob_tokens": [ + -2.977605104446411 + ], + "logprob_reward": -0.10260503739118576, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.10260503739118576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28857624530792236, + 0.04706902056932449, + 0.02019495517015457, + 0.13456889986991882, + 0.13751406967639923, + 0.06357062608003616, + 0.049399279057979584, + 0.06092289462685585, + 0.09610254317522049, + 0.10208146274089813 + ], + "credit_map": [ + 0.021311549469828606, + 0.011920924298465252, + 0.0965857207775116, + 0.038591574877500534, + 0.01684146746993065, + 0.07250078022480011, + 0.014392874203622341, + 0.01643134281039238, + 0.0035411196295171976, + 0.00478322571143508, + 0.0064953407272696495, + 0.0055128731764853, + 0.0035814670845866203, + 0.0035532009787857533, + 0.011594293639063835, + 0.008007493801414967, + 0.0024065012112259865, + 0.001992701319977641, + 0.0021903763990849257, + 0.002223731018602848, + 0.0022424948401749134, + 0.0036211698316037655, + 0.002118052216246724, + 0.003399924375116825, + 0.011157873086631298, + 0.015163066796958447, + 0.013515178114175797, + 0.014505471102893353, + 0.022319039329886436, + 0.026278287172317505, + 0.015970677137374878, + 0.015659300610423088, + 0.010878004133701324, + 0.012608286924660206, + 0.009920359589159489, + 0.012858704663813114, + 0.017206378281116486, + 0.04450644552707672, + 0.008973600342869759, + 0.02056228183209896, + 0.00898753758519888, + 0.005882903002202511, + 0.008337905630469322, + 0.007253976538777351, + 0.008349902927875519, + 0.008996812626719475, + 0.008043159730732441, + 0.007718421518802643, + 0.0034439803566783667, + 0.004734932444989681, + 0.011849743314087391, + 0.007642943877726793, + 0.005727877374738455, + 0.007381671108305454, + 0.004040891770273447, + 0.00457723205909133, + 0.006701335310935974, + 0.006057186983525753, + 0.009784323163330555, + 0.008521880023181438, + 0.006046943366527557, + 0.009422427043318748, + 0.006785680074244738, + 0.0076031130738556385, + 0.004865099210292101, + 0.005227650981396437, + 0.028585512191057205, + 0.003954616840928793, + 0.007176016923040152, + 0.03805474564433098, + 0.005034112371504307, + 0.003204776905477047, + 0.007478879764676094, + 0.004412279464304447, + 0.007094547618180513, + 0.0048184567131102085, + 0.00575066264718771, + 0.03561588004231453, + 0.005995298735797405, + 0.03091544844210148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.10260503739118576 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486039936542511 + ], + "opposite_logprob_tokens": [ + -1.2236039638519287 + ], + "logprob_reward": -0.3486039936542511, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486039936542511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.299306720495224, + 0.06758733093738556, + 0.0997888371348381, + 0.13205644488334656, + 0.051691845059394836, + 0.046358563005924225, + 0.03771704435348511, + 0.04169345274567604, + 0.0806034505367279, + 0.1431962251663208 + ], + "credit_map": [ + 0.015791339799761772, + 0.050002772361040115, + 0.021413614973425865, + 0.010926992632448673, + 0.018140312284231186, + 0.11247357726097107, + 0.015784693881869316, + 0.05477344989776611, + 0.009213220328092575, + 0.004894852172583342, + 0.008067768067121506, + 0.007129146251827478, + 0.008804483339190483, + 0.005400652065873146, + 0.013441269285976887, + 0.010635941289365292, + 0.013878433033823967, + 0.008575614541769028, + 0.012027030810713768, + 0.014015978202223778, + 0.011237824335694313, + 0.011829535476863384, + 0.01593511551618576, + 0.012289317324757576, + 0.00899070780724287, + 0.01635618880391121, + 0.014730170369148254, + 0.014454823918640614, + 0.02213004045188427, + 0.025361282750964165, + 0.016298510134220123, + 0.013734748587012291, + 0.0031395230907946825, + 0.0021509232465177774, + 0.013536921702325344, + 0.004528747871518135, + 0.0038756723515689373, + 0.009370359592139721, + 0.004539216868579388, + 0.010550484992563725, + 0.006118209101259708, + 0.0059972964227199554, + 0.005475612357258797, + 0.005401201080530882, + 0.007155572064220905, + 0.005390816368162632, + 0.006312619429081678, + 0.0045072417706251144, + 0.002809588797390461, + 0.004860305227339268, + 0.006769907660782337, + 0.006770487874746323, + 0.0036700572818517685, + 0.0058896648697555065, + 0.0036682847421616316, + 0.0032787518575787544, + 0.003060649847611785, + 0.0045617916621267796, + 0.006625139620155096, + 0.004061335697770119, + 0.004565914627164602, + 0.004261761903762817, + 0.004526367876678705, + 0.010030495934188366, + 0.003909109625965357, + 0.005003653932362795, + 0.022622089833021164, + 0.003060358576476574, + 0.005868010222911835, + 0.03165674954652786, + 0.00507737323641777, + 0.0034061188343912363, + 0.013657801784574986, + 0.00632658414542675, + 0.009108095429837704, + 0.018267182633280754, + 0.007750372402369976, + 0.03371211513876915, + 0.008497359231114388, + 0.04587673768401146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.3486039936542511 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819257251918316 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.01819257251918316, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819257251918316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5015062689781189, + 0.07469132542610168, + 0.013638854026794434, + 0.09707271307706833, + 0.21255582571029663, + 0.02055678702890873, + 0.01707722246646881, + 0.02285110391676426, + 0.022767826914787292, + 0.01728212460875511 + ], + "credit_map": [ + 0.03567701205611229, + 0.0618104413151741, + 0.1624988168478012, + 0.06655140966176987, + 0.028882285580039024, + 0.09143108874559402, + 0.02520112879574299, + 0.029454099014401436, + 0.015746114775538445, + 0.008006087504327297, + 0.010278568603098392, + 0.00860854983329773, + 0.007847760803997517, + 0.009052843786776066, + 0.007317728828638792, + 0.007833668030798435, + 0.0017246264033019543, + 0.0015958005096763372, + 0.0014342817012220621, + 0.0014758585020899773, + 0.001820767647586763, + 0.002008490264415741, + 0.0017120449338108301, + 0.0018669844139367342, + 0.010895416140556335, + 0.008348180912435055, + 0.014771661721169949, + 0.014028279110789299, + 0.012949478812515736, + 0.012018851935863495, + 0.01279433537274599, + 0.01126650720834732, + 0.015651443973183632, + 0.017108315601944923, + 0.02730335295200348, + 0.013384551741182804, + 0.020879318937659264, + 0.0837785080075264, + 0.01158685889095068, + 0.0228634774684906, + 0.002605705987662077, + 0.002378163393586874, + 0.002674850868061185, + 0.0027100772131234407, + 0.0023393870797008276, + 0.00321608386002481, + 0.0024944220203906298, + 0.0021380966063588858, + 0.0011161437723785639, + 0.0013343236641958356, + 0.0038856416940689087, + 0.0022886660881340504, + 0.0021814953070133924, + 0.0031481883488595486, + 0.0012318069348111749, + 0.0018909566570073366, + 0.002179614966735244, + 0.0027095144614577293, + 0.0037514809519052505, + 0.002431053202599287, + 0.003499722806736827, + 0.002320632804185152, + 0.00277339038439095, + 0.003185692708939314, + 0.0008734227158129215, + 0.0010576894273981452, + 0.007470839191228151, + 0.0010017866734415293, + 0.001684183836914599, + 0.00902683474123478, + 0.0008737016469240189, + 0.0007793674012646079, + 0.001774900360032916, + 0.0009339055395685136, + 0.0014963975409045815, + 0.0027511317748576403, + 0.0010867550736293197, + 0.006211128085851669, + 0.0009664861718192697, + 0.0020614201202988625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.01819257251918316 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329447865486145 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.02329447865486145, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329447865486145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3632402718067169, + 0.034450553357601166, + 0.02352031134068966, + 0.10863935202360153, + 0.24659480154514313, + 0.05260004848241806, + 0.011684284545481205, + 0.07465749979019165, + 0.03709137439727783, + 0.04752152040600777 + ], + "credit_map": [ + 0.025300057604908943, + 0.016438987106084824, + 0.10773742198944092, + 0.04238613322377205, + 0.02121736854314804, + 0.10767790675163269, + 0.0187352504581213, + 0.023747144266963005, + 0.004004056565463543, + 0.003995623905211687, + 0.004844735376536846, + 0.00372220017015934, + 0.00442439503967762, + 0.003045866498723626, + 0.0043576681055128574, + 0.006056011654436588, + 0.0028523304499685764, + 0.0023030834272503853, + 0.0029255130793899298, + 0.00384535058401525, + 0.002510672900825739, + 0.002845635637640953, + 0.0032483593095093966, + 0.0029893643222749233, + 0.01285967230796814, + 0.0098424619063735, + 0.012254985049366951, + 0.014555304311215878, + 0.014225722290575504, + 0.014974649995565414, + 0.015325509943068027, + 0.014601040631532669, + 0.017827145755290985, + 0.015346542932093143, + 0.04705405235290527, + 0.01178335677832365, + 0.019312862306833267, + 0.10475148260593414, + 0.00973593071103096, + 0.02078341506421566, + 0.006248269695788622, + 0.008338532410562038, + 0.007057263050228357, + 0.006925716996192932, + 0.006463584955781698, + 0.0067553166300058365, + 0.005612451117485762, + 0.005198912229388952, + 0.0007185800350271165, + 0.001157892169430852, + 0.0025017312727868557, + 0.0019083104562014341, + 0.0011306408559903502, + 0.0019912617281079292, + 0.0011766711249947548, + 0.0010991967283189297, + 0.004994491580873728, + 0.006588623858988285, + 0.011539202183485031, + 0.00731582660228014, + 0.009871317073702812, + 0.009049364365637302, + 0.008563484996557236, + 0.01673518307507038, + 0.0027906266041100025, + 0.002265388146042824, + 0.010067141614854336, + 0.002006887923926115, + 0.0028244336135685444, + 0.013007874600589275, + 0.0024037035182118416, + 0.0017253191908821464, + 0.0036523304879665375, + 0.002287117298692465, + 0.0032843006774783134, + 0.0025773439556360245, + 0.0028213809709995985, + 0.017592614516615868, + 0.0029846616089344025, + 0.012321769259870052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.02329447865486145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863579474389553 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.00863579474389553, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863579474389553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4529460072517395, + 0.11577299982309341, + 0.04484400153160095, + 0.1318361908197403, + 0.1560358852148056, + 0.030338004231452942, + 0.01320953480899334, + 0.01107662171125412, + 0.015234792605042458, + 0.028705868870019913 + ], + "credit_map": [ + 0.025598758831620216, + 0.10089319944381714, + 0.06311529129743576, + 0.02410309948027134, + 0.0250100027769804, + 0.15800660848617554, + 0.019825724884867668, + 0.03639335557818413, + 0.01830662228167057, + 0.018011437729001045, + 0.013595263473689556, + 0.012490295805037022, + 0.012617502361536026, + 0.017741139978170395, + 0.011914054863154888, + 0.011096692644059658, + 0.006591477431356907, + 0.0039459941908717155, + 0.004819360561668873, + 0.005819015670567751, + 0.004561282228678465, + 0.005603723227977753, + 0.007261439226567745, + 0.006241710390895605, + 0.014596781693398952, + 0.014093837700784206, + 0.014637908898293972, + 0.019289959222078323, + 0.01514297816902399, + 0.015236693434417248, + 0.02110559120774269, + 0.01773245632648468, + 0.011272894218564034, + 0.008836517110466957, + 0.035450760275125504, + 0.006664217449724674, + 0.011294947005808353, + 0.06447909772396088, + 0.005724711809307337, + 0.012312774546444416, + 0.0026382480282336473, + 0.0030033080838620663, + 0.004044593311846256, + 0.004428083077073097, + 0.004132271744310856, + 0.003888301784172654, + 0.004943125881254673, + 0.003260075580328703, + 0.0010651734191924334, + 0.0012233532033860683, + 0.0023958932142704725, + 0.0019553150050342083, + 0.0014469679445028305, + 0.0023370853159576654, + 0.0013948363484814763, + 0.0013909112894907594, + 0.0017886171117424965, + 0.0015378928510472178, + 0.0015383291756734252, + 0.0011264409404248, + 0.0013832170516252518, + 0.001292708213441074, + 0.0012534849811345339, + 0.00115593196824193, + 0.0018361617112532258, + 0.0010389972012490034, + 0.004132718313485384, + 0.0008246619836427271, + 0.0014545454178005457, + 0.004251271951943636, + 0.0008949907496571541, + 0.0008014461491256952, + 0.003243798855692148, + 0.0017390992725268006, + 0.002475419081747532, + 0.005931015592068434, + 0.0018288377905264497, + 0.009696158580482006, + 0.0016500658821314573, + 0.0021414763759821653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.00863579474389553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long shot, slowly pulling back, captures a breathtaking scene inside an Olympic-standard indoor swimming pool. At the center of the frame, a muscular male swimmer, clad in a white swim cap, black goggles, and tight-fitting black racing trunks, surges through the clear, azure water with an almost ferocious force. Spotlights from the gymnasium's ceiling cast bright, sharp beams of light across the surface, creating a constantly shifting, distorted play of light and shadow beneath the turbulent water. Each of his arm strokes slices through the water with incredible force, seemingly ripping through the very surface, sending up massive white splashes and turbulent currents. Under the surface, countless tiny bubbles, stirred up by his powerful leg kicks, form tumbling white trails, creating the visual effect of a violent submarine landslide. The immense force of this impact causes the entire pool's surface to vibrate violently, waves radiating from him crashing against the edge, and the clear black lane lines distort and sway beneath the surface. The entire image is filled with the tension of power, speed and extreme concentration. The contrast between the deep blue and the dazzling white foam perfectly captures the breathtaking epic feeling and tense atmosphere of the Olympic final.", + "seed": 25369, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_2/index_8280_seed_25369_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09685040786862373, + "vlm_logprob": -0.09685040786862373, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.09685040786862373, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1421964168548584 + ], + "opposite_logprob_tokens": [ + -2.5171964168548584 + ], + "logprob_reward": -0.1421964168548584, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1421964168548584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1710350662469864, + 0.07750988006591797, + 0.2775454521179199, + 0.035236913710832596, + 0.05676294490695, + 0.043004944920539856, + 0.07272559404373169, + 0.19568444788455963, + 0.030914483591914177, + 0.03958028927445412 + ], + "credit_map": [ + 0.02226167917251587, + 0.01862902007997036, + 0.0194073673337698, + 0.01821894571185112, + 0.020879890769720078, + 0.023026222363114357, + 0.026773877441883087, + 0.021838044747710228, + 0.007743057794868946, + 0.006134373601526022, + 0.011113679967820644, + 0.009347239509224892, + 0.009114251472055912, + 0.007767121307551861, + 0.01697470247745514, + 0.009315447881817818, + 0.02625652216374874, + 0.019243771210312843, + 0.06783472746610641, + 0.024016443639993668, + 0.0838688537478447, + 0.015863237902522087, + 0.01631901226937771, + 0.02414288930594921, + 0.0034696513321250677, + 0.0028235481586307287, + 0.00444047199562192, + 0.0037662803661078215, + 0.0027781506069004536, + 0.003088917350396514, + 0.008218195289373398, + 0.00665169907733798, + 0.015182903036475182, + 0.005802625324577093, + 0.007181321736425161, + 0.005374023225158453, + 0.005253864452242851, + 0.006005545146763325, + 0.004937265999615192, + 0.007025389466434717, + 0.003932534251362085, + 0.004919317550957203, + 0.005094450898468494, + 0.007838379591703415, + 0.005884926300495863, + 0.004832901991903782, + 0.004942038096487522, + 0.00556039297953248, + 0.0050333100371062756, + 0.005478843580931425, + 0.00859589409083128, + 0.007770271506160498, + 0.012034067884087563, + 0.011800234206020832, + 0.014747543260455132, + 0.0072654238902032375, + 0.019413474947214127, + 0.019938087090849876, + 0.020181508734822273, + 0.03381822630763054, + 0.02657930739223957, + 0.03190452232956886, + 0.02392232045531273, + 0.019926970824599266, + 0.003088277531787753, + 0.003575695212930441, + 0.0043305144645273685, + 0.0045233117416501045, + 0.003847791813313961, + 0.004766383673995733, + 0.002535988809540868, + 0.0042465198785066605, + 0.003966132644563913, + 0.0038359351456165314, + 0.006649497896432877, + 0.009446801617741585, + 0.003494173288345337, + 0.0036463653668761253, + 0.004718668758869171, + 0.003822709433734417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.1421964168548584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25208163261413574 + ], + "opposite_logprob_tokens": [ + -1.5020816326141357 + ], + "logprob_reward": -0.25208163261413574, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25208163261413574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26862671971321106, + 0.031061261892318726, + 0.16372884809970856, + 0.05866491422057152, + 0.04472486302256584, + 0.0685340166091919, + 0.09389884769916534, + 0.08958935737609863, + 0.10788065940141678, + 0.07329045981168747 + ], + "credit_map": [ + 0.04076472297310829, + 0.01895851269364357, + 0.03397892415523529, + 0.02581389620900154, + 0.027794022113084793, + 0.050269003957509995, + 0.03462102636694908, + 0.03642662614583969, + 0.0036645301152020693, + 0.002764082746580243, + 0.004197250120341778, + 0.004762142430990934, + 0.004850161727517843, + 0.002826689276844263, + 0.004192756488919258, + 0.0038036489859223366, + 0.013226382434368134, + 0.012649684213101864, + 0.018463268876075745, + 0.01609569415450096, + 0.058972883969545364, + 0.012797356583178043, + 0.01283233892172575, + 0.018691236153244972, + 0.0066376132890582085, + 0.005764923989772797, + 0.008010070770978928, + 0.008670459501445293, + 0.005705675575882196, + 0.007243425119668245, + 0.009585240855813026, + 0.007047504186630249, + 0.008929215371608734, + 0.004283773712813854, + 0.005920386873185635, + 0.004136549308896065, + 0.004006807692348957, + 0.004814981948584318, + 0.0057123019360005856, + 0.0069208452478051186, + 0.0072154030203819275, + 0.008320184424519539, + 0.008683321997523308, + 0.009550481103360653, + 0.010818719863891602, + 0.007277878001332283, + 0.008943406865000725, + 0.007724623661488295, + 0.007716904394328594, + 0.01624901033937931, + 0.011298180557787418, + 0.010138659738004208, + 0.009333007037639618, + 0.012823658995330334, + 0.01647581160068512, + 0.009863614104688168, + 0.007874878123402596, + 0.014887392520904541, + 0.013131942600011826, + 0.015768717974424362, + 0.00915595330297947, + 0.009964106604456902, + 0.007917515933513641, + 0.010888857766985893, + 0.010888726450502872, + 0.013538077473640442, + 0.015459793619811535, + 0.015158827416598797, + 0.015025036409497261, + 0.01050966139882803, + 0.010212700814008713, + 0.017087837681174278, + 0.007227239664644003, + 0.008344151079654694, + 0.011056790128350258, + 0.015522814355790615, + 0.007554442156106234, + 0.008338198065757751, + 0.007962198927998543, + 0.007284624967724085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.25208163261413574 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054973479360342026 + ], + "opposite_logprob_tokens": [ + -2.9299733638763428 + ], + "logprob_reward": -0.054973479360342026, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054973479360342026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3199538588523865, + 0.06325899064540863, + 0.08893617242574692, + 0.06889975070953369, + 0.09235919266939163, + 0.02884739264845848, + 0.101539745926857, + 0.0653676763176918, + 0.06261584162712097, + 0.10822141915559769 + ], + "credit_map": [ + 0.04618605598807335, + 0.03828253597021103, + 0.036626316606998444, + 0.030782653018832207, + 0.04349509999155998, + 0.05314839631319046, + 0.03679317981004715, + 0.034639619290828705, + 0.0048907166346907616, + 0.003460733685642481, + 0.009832378476858139, + 0.008731402456760406, + 0.005462351720780134, + 0.0049729123711586, + 0.01819765754044056, + 0.0077108354307711124, + 0.006980160251259804, + 0.005046125967055559, + 0.024230673909187317, + 0.007773843593895435, + 0.026851512491703033, + 0.004640680737793446, + 0.0049795289523899555, + 0.008433641865849495, + 0.008824750781059265, + 0.004772672895342112, + 0.01389333512634039, + 0.007669379003345966, + 0.005884896498173475, + 0.0060785687528550625, + 0.00803706981241703, + 0.01373907458037138, + 0.020851515233516693, + 0.009038945659995079, + 0.012327749282121658, + 0.010091329924762249, + 0.007963373325765133, + 0.011125149205327034, + 0.008507812395691872, + 0.012453309260308743, + 0.0019116445910185575, + 0.003698976943269372, + 0.0035758293233811855, + 0.0048215813003480434, + 0.004004413727670908, + 0.0031647919677197933, + 0.0037575310561805964, + 0.003912623040378094, + 0.010554912500083447, + 0.014367855153977871, + 0.01157294400036335, + 0.009924141690135002, + 0.014267027378082275, + 0.01680971495807171, + 0.012593132443726063, + 0.011450017802417278, + 0.005685319658368826, + 0.00835877750068903, + 0.0075658331625163555, + 0.011009595356881618, + 0.006223606877028942, + 0.008949615061283112, + 0.00931776687502861, + 0.008257166482508183, + 0.007663210853934288, + 0.004524685442447662, + 0.008458917029201984, + 0.009262338280677795, + 0.007320288568735123, + 0.006427297834306955, + 0.0091678686439991, + 0.009791233576834202, + 0.009286768734455109, + 0.010743675753474236, + 0.015778370201587677, + 0.01939460076391697, + 0.011178729124367237, + 0.010191376321017742, + 0.01724359206855297, + 0.014404308050870895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.054973479360342026 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361288502812386 + ], + "opposite_logprob_tokens": [ + -3.6513612270355225 + ], + "logprob_reward": -0.026361288502812386, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361288502812386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09513933956623077, + 0.09347428381443024, + 0.312686562538147, + 0.022394316270947456, + 0.07019159942865372, + 0.09548846632242203, + 0.1739378124475479, + 0.039333127439022064, + 0.055505067110061646, + 0.041849397122859955 + ], + "credit_map": [ + 0.01517447642982006, + 0.009475202299654484, + 0.011469537392258644, + 0.009324685670435429, + 0.0112704923376441, + 0.015648605301976204, + 0.01308201439678669, + 0.009694328531622887, + 0.011166395619511604, + 0.006979405879974365, + 0.01064817514270544, + 0.011474265716969967, + 0.016450194641947746, + 0.008371105417609215, + 0.016399752348661423, + 0.011984989047050476, + 0.027298439294099808, + 0.021167315542697906, + 0.07633522152900696, + 0.027915678918361664, + 0.09901535511016846, + 0.01909106969833374, + 0.017865175381302834, + 0.023998301476240158, + 0.003246689448133111, + 0.0022012456320226192, + 0.0035233027301728725, + 0.0028051710687577724, + 0.0025307636242359877, + 0.0021834608633071184, + 0.002903222106397152, + 0.0030004612635821104, + 0.015041541308164597, + 0.006833494175225496, + 0.008852786384522915, + 0.005436474457383156, + 0.007483207620680332, + 0.008713369257748127, + 0.008626067079603672, + 0.009204655885696411, + 0.010708298534154892, + 0.012065129354596138, + 0.011601896025240421, + 0.014536061324179173, + 0.011824870482087135, + 0.013149270787835121, + 0.010432591661810875, + 0.011170350946485996, + 0.015211055986583233, + 0.03297225385904312, + 0.019989479333162308, + 0.025175979360938072, + 0.015985634177923203, + 0.026504142209887505, + 0.016518060117959976, + 0.021581213921308517, + 0.004226823337376118, + 0.0062467120587825775, + 0.004414012189954519, + 0.006037883460521698, + 0.003823778824880719, + 0.006632765755057335, + 0.00320400670170784, + 0.004747144877910614, + 0.00584445521235466, + 0.00499509135261178, + 0.006888180039823055, + 0.008874147199094296, + 0.00746379466727376, + 0.007035182788968086, + 0.006751720793545246, + 0.0076524908654391766, + 0.004338740836828947, + 0.0039702909998595715, + 0.0056527298875153065, + 0.0059765237383544445, + 0.00511213717982173, + 0.0044273752719163895, + 0.004827989265322685, + 0.007543609011918306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.026361288502812386 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639222010970116 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.008639222010970116, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639222010970116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1488906592130661, + 0.06396462023258209, + 0.23747245967388153, + 0.29774728417396545, + 0.046707455068826675, + 0.037481892853975296, + 0.05335152894258499, + 0.04515682905912399, + 0.020006472244858742, + 0.049220867455005646 + ], + "credit_map": [ + 0.017836593091487885, + 0.010729106143116951, + 0.01976470649242401, + 0.015687106177210808, + 0.018337957561016083, + 0.027250828221440315, + 0.02029409445822239, + 0.01899026148021221, + 0.0073476675897836685, + 0.005666062235832214, + 0.00766475684940815, + 0.010443377308547497, + 0.008375492878258228, + 0.006090590264648199, + 0.010522962547838688, + 0.007853715680539608, + 0.018874024972319603, + 0.01741238683462143, + 0.05353207886219025, + 0.01949136145412922, + 0.0872410237789154, + 0.01643148809671402, + 0.011701222509145737, + 0.012788882479071617, + 0.03073710948228836, + 0.02011759951710701, + 0.042900100350379944, + 0.034949444234371185, + 0.03010581061244011, + 0.022990861907601357, + 0.06108444184064865, + 0.05486191064119339, + 0.009307765401899815, + 0.00470880838111043, + 0.006760957185178995, + 0.0060651786625385284, + 0.003679585875943303, + 0.00534056918695569, + 0.0037702496629208326, + 0.007074336986988783, + 0.0035605067387223244, + 0.004775542765855789, + 0.004625541158020496, + 0.0054737962782382965, + 0.004556574858725071, + 0.005014814902096987, + 0.0051949298940598965, + 0.0042801834642887115, + 0.00572493951767683, + 0.00711190328001976, + 0.006781098898500204, + 0.005334611050784588, + 0.006829987745732069, + 0.006516580004245043, + 0.007665816694498062, + 0.007386595010757446, + 0.004623293410986662, + 0.006626757327467203, + 0.005458889529109001, + 0.007931026630103588, + 0.003950721118599176, + 0.006233841180801392, + 0.004766878206282854, + 0.005565422121435404, + 0.0024518040008842945, + 0.002023711334913969, + 0.0025640225503593683, + 0.003354900749400258, + 0.002433652989566326, + 0.002471810206770897, + 0.0019025250803679228, + 0.0028040448669344187, + 0.00425243005156517, + 0.005940599367022514, + 0.0071232994087040424, + 0.008610470220446587, + 0.004892767407000065, + 0.0063094221986830235, + 0.006860930472612381, + 0.005230948328971863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.008639222010970116 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, long shot, slowly pulling back, captures a breathtaking scene inside an Olympic-standard indoor swimming pool. At the center of the frame, a muscular male swimmer, clad in a white swim cap, black goggles, and tight-fitting black racing trunks, surges through the clear, azure water with an almost ferocious force. Spotlights from the gymnasium's ceiling cast bright, sharp beams of light across the surface, creating a constantly shifting, distorted play of light and shadow beneath the turbulent water. Each of his arm strokes slices through the water with incredible force, seemingly ripping through the very surface, sending up massive white splashes and turbulent currents. Under the surface, countless tiny bubbles, stirred up by his powerful leg kicks, form tumbling white trails, creating the visual effect of a violent submarine landslide. The immense force of this impact causes the entire pool's surface to vibrate violently, waves radiating from him crashing against the edge, and the clear black lane lines distort and sway beneath the surface. The entire image is filled with the tension of power, speed and extreme concentration. The contrast between the deep blue and the dazzling white foam perfectly captures the breathtaking epic feeling and tense atmosphere of the Olympic final.", + "seed": 25370, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_2/index_8280_seed_25370_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15077900271862746, + "vlm_logprob": -0.15077900271862746, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.15077900271862746, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.29264694452285767 + ], + "opposite_logprob_tokens": [ + -1.792646884918213 + ], + "logprob_reward": -0.29264694452285767, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.29264694452285767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058206792920827866, + 0.3709985315799713, + 0.07548455148935318, + 0.11223440617322922, + 0.0835355892777443, + 0.020616192370653152, + 0.07339952886104584, + 0.04667333513498306, + 0.12212523072957993, + 0.036725834012031555 + ], + "credit_map": [ + 0.005728152114897966, + 0.00612901896238327, + 0.009127908386290073, + 0.006525603588670492, + 0.005974640138447285, + 0.007122039329260588, + 0.00751852011308074, + 0.010080913081765175, + 0.021106859669089317, + 0.0303146094083786, + 0.06022334843873978, + 0.018500104546546936, + 0.06623056530952454, + 0.042015232145786285, + 0.03474435210227966, + 0.09786351770162582, + 0.006521712522953749, + 0.008527767844498158, + 0.00972194503992796, + 0.009159505367279053, + 0.008155640214681625, + 0.011303271166980267, + 0.011435287073254585, + 0.010659429244697094, + 0.009773977100849152, + 0.011580650694668293, + 0.01855410635471344, + 0.0183627437800169, + 0.008954105898737907, + 0.011989295482635498, + 0.020023740828037262, + 0.01299579069018364, + 0.0034435398411005735, + 0.0036300027277320623, + 0.01812148466706276, + 0.018211325630545616, + 0.003375317668542266, + 0.0034688683226704597, + 0.013875887729227543, + 0.019409166648983955, + 0.001590872765518725, + 0.0023082972038537264, + 0.002784560900181532, + 0.0028762042056769133, + 0.0032276033889502287, + 0.0037524246145039797, + 0.0019358140416443348, + 0.002140415832400322, + 0.007008471060544252, + 0.006449834909290075, + 0.019549660384655, + 0.018665874376893044, + 0.005625133868306875, + 0.004007479175925255, + 0.006422810722142458, + 0.0056702629663050175, + 0.004410189110785723, + 0.006394148338586092, + 0.006562316790223122, + 0.005730781704187393, + 0.008924964815378189, + 0.004622924607247114, + 0.005337960552424192, + 0.0046900492161512375, + 0.009882737882435322, + 0.007568187080323696, + 0.029401596635580063, + 0.018359048292040825, + 0.010661473497748375, + 0.01414399966597557, + 0.013494505546987057, + 0.01861369051039219, + 0.0017743041971698403, + 0.001868451596237719, + 0.0024173473939299583, + 0.01276780292391777, + 0.001880402909591794, + 0.001535473158583045, + 0.011683108285069466, + 0.0027989454101771116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 3 + ], + "selected_reward": -0.29264694452285767 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134554326534271 + ], + "opposite_logprob_tokens": [ + -1.3134554624557495 + ], + "logprob_reward": -0.3134554326534271, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134554326534271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10188379138708115, + 0.2657393515110016, + 0.07238177955150604, + 0.10461096465587616, + 0.08951470255851746, + 0.10416179150342941, + 0.12317366898059845, + 0.028483381494879723, + 0.04926718771457672, + 0.060783375054597855 + ], + "credit_map": [ + 0.011568145826458931, + 0.011235319077968597, + 0.013487199321389198, + 0.016142938286066055, + 0.015276714228093624, + 0.010769125074148178, + 0.010584671050310135, + 0.012819684110581875, + 0.014873666688799858, + 0.02199625037610531, + 0.045713238418102264, + 0.015667548403143883, + 0.044146306812763214, + 0.02897973731160164, + 0.023151911795139313, + 0.07121068239212036, + 0.006021869834512472, + 0.00804171897470951, + 0.008806215599179268, + 0.006754712201654911, + 0.009144581854343414, + 0.01170183252543211, + 0.011249260976910591, + 0.010661585256457329, + 0.007734993938356638, + 0.011001401580870152, + 0.014878797344863415, + 0.01970052719116211, + 0.010535378940403461, + 0.010289830155670643, + 0.015920182690024376, + 0.014549847692251205, + 0.009041687473654747, + 0.009176519699394703, + 0.014492410235106945, + 0.01342127937823534, + 0.00739167770370841, + 0.005843117833137512, + 0.01690998300909996, + 0.013238026760518551, + 0.013347637839615345, + 0.01795903779566288, + 0.01241723820567131, + 0.018950937315821648, + 0.0134115070104599, + 0.010832834988832474, + 0.009101297706365585, + 0.00814130064100027, + 0.009299245662987232, + 0.0126732736825943, + 0.027022799476981163, + 0.026906471699476242, + 0.016211288049817085, + 0.011096016503870487, + 0.009163038805127144, + 0.010801532305777073, + 0.003587977960705757, + 0.0036566187627613544, + 0.0049337297677993774, + 0.0040959930047392845, + 0.003488075453788042, + 0.0022704817820340395, + 0.00338260387070477, + 0.0030679013580083847, + 0.003795877331867814, + 0.0049091726541519165, + 0.005715919192880392, + 0.006327622104436159, + 0.009431177750229836, + 0.0041758171282708645, + 0.00882961880415678, + 0.006081984844058752, + 0.0024310259614139795, + 0.0027606841176748276, + 0.0048105656169354916, + 0.024595653638243675, + 0.002353994408622384, + 0.0018530874513089657, + 0.01994619518518448, + 0.0020321670453995466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.3134554326534271 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300592124462128 + ], + "opposite_logprob_tokens": [ + -2.2380058765411377 + ], + "logprob_reward": -0.11300592124462128, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300592124462128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0380348339676857, + 0.44081348180770874, + 0.1278369277715683, + 0.06631492078304291, + 0.015754593536257744, + 0.07374849170446396, + 0.035334616899490356, + 0.06273104250431061, + 0.027791529893875122, + 0.11163955181837082 + ], + "credit_map": [ + 0.003961934242397547, + 0.0042255716398358345, + 0.00477969367057085, + 0.0037091695703566074, + 0.0060329921543598175, + 0.006043077912181616, + 0.004795515909790993, + 0.004486879333853722, + 0.025359204038977623, + 0.03551572188735008, + 0.06841219961643219, + 0.021511387079954147, + 0.07971825450658798, + 0.052521415054798126, + 0.04224910959601402, + 0.11552617698907852, + 0.008111433126032352, + 0.01075026299804449, + 0.017252692952752113, + 0.016965249553322792, + 0.011895349249243736, + 0.01655529998242855, + 0.019896941259503365, + 0.026409689337015152, + 0.007799526210874319, + 0.006664975080639124, + 0.011142786592245102, + 0.010957074351608753, + 0.005249237176030874, + 0.007778226863592863, + 0.008563278242945671, + 0.008159816265106201, + 0.0017953445203602314, + 0.001843304606154561, + 0.0020670886151492596, + 0.0026630174834281206, + 0.0012335545616224408, + 0.0012117807054892182, + 0.002105989493429661, + 0.0028345121536403894, + 0.005478912964463234, + 0.0064195552840828896, + 0.012836947105824947, + 0.01997198350727558, + 0.0058335005305707455, + 0.007533928379416466, + 0.009673378430306911, + 0.0060002775862813, + 0.004136072471737862, + 0.003696645377203822, + 0.0064560831524431705, + 0.005584184546023607, + 0.003722689114511013, + 0.0034912421833723783, + 0.004101433791220188, + 0.004146264865994453, + 0.009247600100934505, + 0.008269138634204865, + 0.009089303202927113, + 0.008035066537559032, + 0.006377516780048609, + 0.0063870251178741455, + 0.00889669917523861, + 0.00642869109287858, + 0.002725625177845359, + 0.002196624642238021, + 0.003999690525233746, + 0.0036510592326521873, + 0.003080088645219803, + 0.0031437883153557777, + 0.0037941515911370516, + 0.005200497806072235, + 0.004126599058508873, + 0.00683974102139473, + 0.0071166385896503925, + 0.038668714463710785, + 0.007092067040503025, + 0.003799386555328965, + 0.03874301537871361, + 0.0052533806301653385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.11300592124462128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591532811522484 + ], + "opposite_logprob_tokens": [ + -3.8955914974212646 + ], + "logprob_reward": -0.020591532811522484, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591532811522484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05732516571879387, + 0.5546503067016602, + 0.03562946990132332, + 0.03539228066802025, + 0.031680114567279816, + 0.05096617713570595, + 0.059598810970783234, + 0.064960777759552, + 0.057608142495155334, + 0.052188754081726074 + ], + "credit_map": [ + 0.006822461728006601, + 0.00840337947010994, + 0.008889997377991676, + 0.007033040747046471, + 0.006069370545446873, + 0.00601885374635458, + 0.007242744788527489, + 0.006845314521342516, + 0.030841894447803497, + 0.044122714549303055, + 0.09706062078475952, + 0.027996232733130455, + 0.09428879618644714, + 0.059380050748586655, + 0.05138289928436279, + 0.1495770812034607, + 0.002765863901004195, + 0.004522509407252073, + 0.004632546566426754, + 0.004816032480448484, + 0.0033641860354691744, + 0.0036584048066288233, + 0.006099681835621595, + 0.0057702455669641495, + 0.002359643578529358, + 0.004193092230707407, + 0.004480530042201281, + 0.005246541928499937, + 0.003972755279392004, + 0.0038745335768908262, + 0.005044270772486925, + 0.006220912095159292, + 0.0023294154088944197, + 0.002028741640970111, + 0.008504213765263557, + 0.008246510289609432, + 0.0017992990324273705, + 0.001883986871689558, + 0.0037247755099087954, + 0.003163173794746399, + 0.003354971529915929, + 0.004097247030586004, + 0.0076013957150280476, + 0.014217421412467957, + 0.006365164648741484, + 0.006204943172633648, + 0.005400801543146372, + 0.0037242339458316565, + 0.004554301500320435, + 0.006943147163838148, + 0.010976850986480713, + 0.012122919782996178, + 0.007115004118531942, + 0.005508982110768557, + 0.00532190129160881, + 0.007055705413222313, + 0.008010102435946465, + 0.008941291831433773, + 0.00892141368240118, + 0.008135645650327206, + 0.0091117974370718, + 0.007533751428127289, + 0.008097128942608833, + 0.006209648679941893, + 0.004280528053641319, + 0.0031299160327762365, + 0.018210794776678085, + 0.012529628351330757, + 0.0030160530004650354, + 0.0038795347791165113, + 0.00546973617747426, + 0.007091955281794071, + 0.002200834220275283, + 0.003598631825298071, + 0.003975707106292248, + 0.017421405762434006, + 0.002962749917060137, + 0.002958194352686405, + 0.015614939853549004, + 0.003456290578469634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.020591532811522484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly pulling back during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set inside an Olympic-standard indoor swimming pool?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the swimmer wearing a white swim cap, black goggles, and tight-fitting black racing trunks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are bright, sharp beams of light from ceiling spotlights visible casting distorted patterns on the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195182360708714 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014195182360708714, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195182360708714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04123714566230774, + 0.25179532170295715, + 0.03927246853709221, + 0.27371805906295776, + 0.26950958371162415, + 0.018821140751242638, + 0.03856236860156059, + 0.009514305740594864, + 0.03684379905462265, + 0.020725762471556664 + ], + "credit_map": [ + 0.00393211655318737, + 0.005148523952811956, + 0.008843076415359974, + 0.004331101663410664, + 0.004731956869363785, + 0.0036199495662003756, + 0.004887544550001621, + 0.0057428739964962006, + 0.014569545164704323, + 0.01874462515115738, + 0.04659045487642288, + 0.012557119131088257, + 0.04068181291222572, + 0.024814575910568237, + 0.023424893617630005, + 0.0704123005270958, + 0.002613465301692486, + 0.003906617406755686, + 0.005367468111217022, + 0.003816581331193447, + 0.004070690833032131, + 0.005797681398689747, + 0.005669361911714077, + 0.008030601777136326, + 0.02570617012679577, + 0.025877146050333977, + 0.04289482161402702, + 0.04261009395122528, + 0.02429824322462082, + 0.026880459859967232, + 0.058732450008392334, + 0.026718683540821075, + 0.0164879709482193, + 0.016948360949754715, + 0.09426857531070709, + 0.08848808705806732, + 0.015061725862324238, + 0.013165685348212719, + 0.012462465092539787, + 0.012626735493540764, + 0.00352504407055676, + 0.0022684019058942795, + 0.0024142039474099874, + 0.0024796517100185156, + 0.0020244577899575233, + 0.0024843369610607624, + 0.0015191588317975402, + 0.002105887746438384, + 0.003496715100482106, + 0.002696406561881304, + 0.009702255018055439, + 0.009237871505320072, + 0.004705233033746481, + 0.003000858938321471, + 0.00300303241237998, + 0.0027199990581721067, + 0.0010156262433156371, + 0.0013374595437198877, + 0.0010235641384497285, + 0.0011226008646190166, + 0.001648244448006153, + 0.0011391225270926952, + 0.001055247033946216, + 0.0011724403593689203, + 0.0024902767036110163, + 0.0016904897056519985, + 0.01142044086009264, + 0.008498376235365868, + 0.002349338959902525, + 0.0033970752265304327, + 0.002908221213147044, + 0.0040895817801356316, + 0.0016301736468449235, + 0.0017262373585253954, + 0.0022453556302934885, + 0.0058166575618088245, + 0.001151116332039237, + 0.0019932787399739027, + 0.004029111936688423, + 0.002133829751983285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.014195182360708714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_4/index_tensor([981, 981])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_4/index_tensor([981, 981])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3c54219716ae371aa26f544057fac71f69462553 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_4/index_tensor([981, 981])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot begins from an extremely low angle, the camera tilting steadily upward to track the billowing black smoke rising violently from the cracked asphalt. Like an ancient beast awakening, this thick, oily column of smoke rises in a chaotic and powerful rhythm, its interior streaked with fiery orange-red sparks that flash through the inky black haze. The edges of the smoke column exhibit a complex, feathery curling structure, twisting and expanding as it spreads toward a gloomy, oppressive gray sky. The ground is littered with indistinct industrial debris and charred debris, hinting at the source of the disaster. The entire scene is devoured by the smoke, with only occasional flashes of fiery orange light emanating from the burning core at the base of the column, providing the sole illumination for this devastating and oppressive scene, creating an atmosphere of apocalyptic grandeur and awe.", + "seed": 26585, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_4/index_981_seed_26585_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1272629036102444, + "vlm_logprob": -0.1272629036102444, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.1272629036102444, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5869388580322266 + ], + "opposite_logprob_tokens": [ + -0.8369388580322266 + ], + "logprob_reward": -0.5869388580322266, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5869388580322266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.210683211684227, + 0.11334770917892456, + 0.09924701601266861, + 0.2379419356584549, + 0.07922419905662537, + 0.033760879188776016, + 0.05832039564847946, + 0.0537971593439579, + 0.024462871253490448, + 0.08921466022729874 + ], + "credit_map": [ + 0.02342512644827366, + 0.030529484152793884, + 0.026520846411585808, + 0.029498614370822906, + 0.026805894449353218, + 0.019796354696154594, + 0.029981018975377083, + 0.02412586659193039, + 0.011116642504930496, + 0.007676862645894289, + 0.03851526603102684, + 0.01053859107196331, + 0.012452954426407814, + 0.006294723134487867, + 0.014170367270708084, + 0.01258230023086071, + 0.0067283837124705315, + 0.012266569770872593, + 0.011161199770867825, + 0.012718702666461468, + 0.011700560338795185, + 0.012108484283089638, + 0.013513178564608097, + 0.019049933180212975, + 0.025914525613188744, + 0.033845026046037674, + 0.03552936017513275, + 0.026079485192894936, + 0.026844056323170662, + 0.024998899549245834, + 0.04300303757190704, + 0.021727539598941803, + 0.0068602473475039005, + 0.013559968210756779, + 0.004261543042957783, + 0.01499374769628048, + 0.011872699484229088, + 0.00887732207775116, + 0.014596867375075817, + 0.004201805219054222, + 0.0028359664138406515, + 0.006723891012370586, + 0.0018362489063292742, + 0.002892171498388052, + 0.005766242276877165, + 0.00852013099938631, + 0.0039860946126282215, + 0.0012001317227259278, + 0.004830693360418081, + 0.004643294028937817, + 0.0037786958273500204, + 0.007376424502581358, + 0.010091286152601242, + 0.009986831806600094, + 0.01407708041369915, + 0.0035360893234610558, + 0.004025986418128014, + 0.006971547845751047, + 0.0058198776096105576, + 0.009472749195992947, + 0.0074172792956233025, + 0.007685012649744749, + 0.0071371071971952915, + 0.00526759959757328, + 0.0017295504221692681, + 0.0028719145338982344, + 0.0015802859561517835, + 0.002964170416817069, + 0.0064863236621022224, + 0.0038823080249130726, + 0.002817226340994239, + 0.0021310909651219845, + 0.0035957912914454937, + 0.004310290329158306, + 0.004619085695594549, + 0.004187299404293299, + 0.010405211709439754, + 0.02981729991734028, + 0.009781061671674252, + 0.02249862812459469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.5869388580322266 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005276799667626619 + ], + "opposite_logprob_tokens": [ + -5.255276679992676 + ], + "logprob_reward": -0.005276799667626619, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005276799667626619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2525928318500519, + 0.15067870914936066, + 0.05708055570721626, + 0.08823669701814651, + 0.11524064838886261, + 0.16029076278209686, + 0.07601208239793777, + 0.037700213491916656, + 0.027643844485282898, + 0.03452359139919281 + ], + "credit_map": [ + 0.02062210440635681, + 0.02544131875038147, + 0.03300003334879875, + 0.04962339252233505, + 0.02723522298038006, + 0.01607624813914299, + 0.03869902715086937, + 0.041895486414432526, + 0.009079008363187313, + 0.009276757948100567, + 0.047218888998031616, + 0.016337255015969276, + 0.017217138782143593, + 0.009946554899215698, + 0.019847825169563293, + 0.021755276247859, + 0.004453858360648155, + 0.008726699277758598, + 0.005858559161424637, + 0.009742526337504387, + 0.006193611305207014, + 0.005152846686542034, + 0.00831546913832426, + 0.008636988699436188, + 0.010911849327385426, + 0.010874406434595585, + 0.0183273758739233, + 0.010111544281244278, + 0.006710255518555641, + 0.007843275554478168, + 0.011348257772624493, + 0.012109731324017048, + 0.011286047287285328, + 0.015767015516757965, + 0.010879963636398315, + 0.01736271381378174, + 0.01837106980383396, + 0.01571054942905903, + 0.01875299960374832, + 0.007110282313078642, + 0.012202841229736805, + 0.022133223712444305, + 0.010248403064906597, + 0.017314674332737923, + 0.02737075835466385, + 0.0316879078745842, + 0.03209878131747246, + 0.007234182208776474, + 0.00897210743278265, + 0.00803417805582285, + 0.005080683622509241, + 0.007221781183034182, + 0.01282318215817213, + 0.012084556743502617, + 0.01055600494146347, + 0.011239586398005486, + 0.002550208242610097, + 0.004526782315224409, + 0.002906200708821416, + 0.005941662006080151, + 0.005927025806158781, + 0.005485984031111002, + 0.005800055339932442, + 0.004562295973300934, + 0.0016128765419125557, + 0.001960759749636054, + 0.00191751669626683, + 0.004419860430061817, + 0.004703432787209749, + 0.004163249395787716, + 0.004344060551375151, + 0.004522085655480623, + 0.0032328623346984386, + 0.0030621227342635393, + 0.003847334301099181, + 0.003843056270852685, + 0.004531968384981155, + 0.006332056596875191, + 0.004202048294246197, + 0.005472142249345779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.005276799667626619 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036206187214702368 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036206187214702368, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036206187214702368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14303654432296753, + 0.07624497264623642, + 0.10207613557577133, + 0.06381428241729736, + 0.14227361977100372, + 0.052624043077230453, + 0.030598876997828484, + 0.01676674745976925, + 0.05436526983976364, + 0.3181995153427124 + ], + "credit_map": [ + 0.00800738949328661, + 0.01217188872396946, + 0.019763346761465073, + 0.04265565425157547, + 0.008422741666436195, + 0.005064816679805517, + 0.030303223058581352, + 0.01664748601615429, + 0.00739094614982605, + 0.008038180880248547, + 0.01080185454338789, + 0.0113057317212224, + 0.009755279868841171, + 0.006868965923786163, + 0.014262263663113117, + 0.007821744307875633, + 0.008496218360960484, + 0.010216401889920235, + 0.012392950244247913, + 0.014906742610037327, + 0.01186357531696558, + 0.006762451026588678, + 0.01879836432635784, + 0.01863943599164486, + 0.0049801236018538475, + 0.004146452061831951, + 0.021228231489658356, + 0.005051208660006523, + 0.004956770222634077, + 0.003643744857981801, + 0.005921858362853527, + 0.013885893858969212, + 0.0140554029494524, + 0.024540096521377563, + 0.008763055317103863, + 0.024483727291226387, + 0.021136479452252388, + 0.022647695615887642, + 0.01908162795007229, + 0.00756553839892149, + 0.0043297940865159035, + 0.010458161123096943, + 0.002783646574243903, + 0.0064384108409285545, + 0.008095191791653633, + 0.009296166710555553, + 0.008949648588895798, + 0.002273024059832096, + 0.0018812932539731264, + 0.0033029126934707165, + 0.00139849993865937, + 0.003157569095492363, + 0.005362064577639103, + 0.007183730602264404, + 0.005764791276305914, + 0.0025480135809630156, + 0.0011538711842149496, + 0.002137667266651988, + 0.0015862893778830767, + 0.0026137575041502714, + 0.0027454744558781385, + 0.0021276799961924553, + 0.002400690456852317, + 0.0020013179164379835, + 0.0040897466242313385, + 0.00705355079844594, + 0.004906145390123129, + 0.009614964947104454, + 0.007587325759232044, + 0.005855646915733814, + 0.008504979312419891, + 0.006752909626811743, + 0.01649685762822628, + 0.01603626273572445, + 0.014280243776738644, + 0.018949024379253387, + 0.029926840215921402, + 0.09213971346616745, + 0.033377815037965775, + 0.09699276834726334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.0036206187214702368 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033740513026714325 + ], + "opposite_logprob_tokens": [ + -3.408740520477295 + ], + "logprob_reward": -0.033740513026714325, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033740513026714325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21852931380271912, + 0.12150438129901886, + 0.12629112601280212, + 0.08666259050369263, + 0.12330621480941772, + 0.08669610321521759, + 0.03526808321475983, + 0.050755564123392105, + 0.0916479080915451, + 0.05933870002627373 + ], + "credit_map": [ + 0.021956615149974823, + 0.023599905893206596, + 0.025628868490457535, + 0.027520382776856422, + 0.02582225762307644, + 0.018756845965981483, + 0.02734268829226494, + 0.04790174961090088, + 0.009172555059194565, + 0.007947932928800583, + 0.03995910659432411, + 0.01205918937921524, + 0.013897374272346497, + 0.007161454763263464, + 0.016881922259926796, + 0.01442483626306057, + 0.01004761178046465, + 0.018320253118872643, + 0.014168310910463333, + 0.02031669206917286, + 0.013789365068078041, + 0.011307885870337486, + 0.016057651489973068, + 0.02228335663676262, + 0.00879113469272852, + 0.009620288386940956, + 0.011474094353616238, + 0.013623636215925217, + 0.008989929221570492, + 0.008669468574225903, + 0.015147790312767029, + 0.01034624595195055, + 0.012396879494190216, + 0.01760299690067768, + 0.0078435018658638, + 0.021820371970534325, + 0.020098470151424408, + 0.013598955236375332, + 0.021711915731430054, + 0.008233118802309036, + 0.008905510418117046, + 0.021166274324059486, + 0.004952055402100086, + 0.008615059778094292, + 0.012491393834352493, + 0.01628449745476246, + 0.009551902301609516, + 0.004729407373815775, + 0.004584749694913626, + 0.0029973401688039303, + 0.002257069805637002, + 0.0029612435027956963, + 0.005630110390484333, + 0.0060285599902272224, + 0.005773186217993498, + 0.005035825539380312, + 0.004130634944885969, + 0.005457883235067129, + 0.004023201297968626, + 0.007822515442967415, + 0.009118744172155857, + 0.006214040797203779, + 0.0064083063043653965, + 0.007580237928777933, + 0.004801461473107338, + 0.005775718484073877, + 0.004584084264934063, + 0.01803349144756794, + 0.014780070632696152, + 0.011646118946373463, + 0.013811134733259678, + 0.018215829506516457, + 0.004607719369232655, + 0.003921409137547016, + 0.0050540887750685215, + 0.006075707264244556, + 0.004696674179285765, + 0.011442914605140686, + 0.006364563014358282, + 0.017175622284412384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.033740513026714325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737728603184223 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737728603184223, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737728603184223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24466702342033386, + 0.07656623423099518, + 0.11692393571138382, + 0.14360587298870087, + 0.09283430874347687, + 0.0487976111471653, + 0.04180225357413292, + 0.027367379516363144, + 0.08409229665994644, + 0.1233430951833725 + ], + "credit_map": [ + 0.01562364213168621, + 0.022481955587863922, + 0.03112155571579933, + 0.06145401671528816, + 0.019209790974855423, + 0.015200111083686352, + 0.041078537702560425, + 0.03849741443991661, + 0.008283779956400394, + 0.006595872342586517, + 0.012439773418009281, + 0.010270936414599419, + 0.008234869688749313, + 0.004787804093211889, + 0.014997999183833599, + 0.010955200530588627, + 0.009545912966132164, + 0.011146372184157372, + 0.015899311751127243, + 0.01304350420832634, + 0.012911507859826088, + 0.011067488230764866, + 0.016566472128033638, + 0.026743371039628983, + 0.013716538436710835, + 0.012391622178256512, + 0.03894165903329849, + 0.012016193941235542, + 0.014110492542386055, + 0.010979985818266869, + 0.01690731570124626, + 0.0245420690625906, + 0.008601240813732147, + 0.013347538188099861, + 0.006116096395999193, + 0.014387893490493298, + 0.015933724120259285, + 0.014476505108177662, + 0.015109570696949959, + 0.004861736670136452, + 0.005231198854744434, + 0.01207191776484251, + 0.0024630443658679724, + 0.005572083871811628, + 0.007476202677935362, + 0.00572692696005106, + 0.006757124327123165, + 0.0034991116262972355, + 0.00521452259272337, + 0.0033016654197126627, + 0.0044046491384506226, + 0.004420313984155655, + 0.006477033253759146, + 0.006329421419650316, + 0.007515476085245609, + 0.004139172378927469, + 0.0024951561354100704, + 0.004622989799827337, + 0.0019354928517714143, + 0.004836989101022482, + 0.003526456421241164, + 0.004296756349503994, + 0.0037097386084496975, + 0.001943801180459559, + 0.006017268169671297, + 0.010655312798917294, + 0.006786760408431292, + 0.016553791239857674, + 0.013134715147316456, + 0.010389997623860836, + 0.010326607152819633, + 0.010227846913039684, + 0.006004160735756159, + 0.008186792954802513, + 0.008563026785850525, + 0.010650386102497578, + 0.014505359344184399, + 0.016897020861506462, + 0.02663121558725834, + 0.03190512955188751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.006737728603184223 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins from an extremely low angle, the camera tilting steadily upward to track the billowing black smoke rising violently from the cracked asphalt. Like an ancient beast awakening, this thick, oily column of smoke rises in a chaotic and powerful rhythm, its interior streaked with fiery orange-red sparks that flash through the inky black haze. The edges of the smoke column exhibit a complex, feathery curling structure, twisting and expanding as it spreads toward a gloomy, oppressive gray sky. The ground is littered with indistinct industrial debris and charred debris, hinting at the source of the disaster. The entire scene is devoured by the smoke, with only occasional flashes of fiery orange light emanating from the burning core at the base of the column, providing the sole illumination for this devastating and oppressive scene, creating an atmosphere of apocalyptic grandeur and awe.", + "seed": 26586, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_4/index_981_seed_26586_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11446750788018108, + "vlm_logprob": -0.11446750788018108, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.11446750788018108, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5325818061828613 + ], + "opposite_logprob_tokens": [ + -0.9075818061828613 + ], + "logprob_reward": -0.5325818061828613, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5325818061828613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22342053055763245, + 0.01729615405201912, + 0.024737263098359108, + 0.07127653062343597, + 0.06319714337587357, + 0.025235790759325027, + 0.05182122811675072, + 0.04295986518263817, + 0.4567130506038666, + 0.023342423141002655 + ], + "credit_map": [ + 0.016120823100209236, + 0.009635270573198795, + 0.07939335703849792, + 0.012742215767502785, + 0.011404972523450851, + 0.06790166348218918, + 0.01151090394705534, + 0.014711329713463783, + 0.001218447694554925, + 0.0016572860768064857, + 0.0039025910664349794, + 0.001271822489798069, + 0.0018337414367124438, + 0.0033045781310647726, + 0.001554755144752562, + 0.002552932361140847, + 0.002359932754188776, + 0.00318039208650589, + 0.0038184337317943573, + 0.003409839002415538, + 0.004140409640967846, + 0.0032354805152863264, + 0.0026454899925738573, + 0.0019472864223644137, + 0.005038800183683634, + 0.022685395553708076, + 0.007961503230035305, + 0.007116974797099829, + 0.013560538180172443, + 0.00514339841902256, + 0.003994547761976719, + 0.005775365978479385, + 0.010086257942020893, + 0.009114931337535381, + 0.012743666768074036, + 0.004576577339321375, + 0.011263716965913773, + 0.006878972519189119, + 0.005750729236751795, + 0.002782289870083332, + 0.0015414227964356542, + 0.0029115304350852966, + 0.00236579030752182, + 0.0032346746884286404, + 0.0030454264488071203, + 0.0030383667908608913, + 0.00504967849701643, + 0.0040489002130925655, + 0.0054368614219129086, + 0.0062814983539283276, + 0.009249086491763592, + 0.005683272611349821, + 0.005462202243506908, + 0.008094405755400658, + 0.006940016057342291, + 0.0046738819219172, + 0.003648653393611312, + 0.004542440641671419, + 0.005991672165691853, + 0.0060835788026452065, + 0.005327952094376087, + 0.006915197242051363, + 0.006223863922059536, + 0.00422650296241045, + 0.048918720334768295, + 0.05603542551398277, + 0.05169444903731346, + 0.054297175258398056, + 0.0939580425620079, + 0.05337214842438698, + 0.07085330784320831, + 0.027583763003349304, + 0.002131642773747444, + 0.002197572262957692, + 0.0023403360974043608, + 0.0026622971054166555, + 0.004933486692607403, + 0.0022377080749720335, + 0.0045895506627857685, + 0.0022498283069580793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.5325818061828613 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00528059434145689 + ], + "opposite_logprob_tokens": [ + -5.255280494689941 + ], + "logprob_reward": -0.00528059434145689, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00528059434145689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6285424828529358, + 0.04244966432452202, + 0.017524559050798416, + 0.03933940455317497, + 0.11317912489175797, + 0.030292930081486702, + 0.027256110683083534, + 0.022215265780687332, + 0.046611446887254715, + 0.03258897364139557 + ], + "credit_map": [ + 0.04801996424794197, + 0.02137651853263378, + 0.24219219386577606, + 0.0340946689248085, + 0.02700529247522354, + 0.19377675652503967, + 0.025191204622387886, + 0.03688589856028557, + 0.0038581774570047855, + 0.0030374464113265276, + 0.00561721995472908, + 0.003708459436893463, + 0.0030791922472417355, + 0.005595691967755556, + 0.004951443988829851, + 0.012602031230926514, + 0.0015635216841474175, + 0.0020955936051905155, + 0.0018254313617944717, + 0.00259943469427526, + 0.003183339722454548, + 0.0022966761607676744, + 0.0018578674644231796, + 0.002102695172652602, + 0.0021210666745901108, + 0.011118405498564243, + 0.003936754073947668, + 0.0031257139053195715, + 0.006649515125900507, + 0.003963491879403591, + 0.0030498187988996506, + 0.005374637432396412, + 0.01007840596139431, + 0.018598856404423714, + 0.02310475893318653, + 0.007248925045132637, + 0.03182831406593323, + 0.009470806457102299, + 0.0074984231032431126, + 0.005350625142455101, + 0.0019464966608211398, + 0.00430048955604434, + 0.002653221134096384, + 0.00329212611541152, + 0.004354685544967651, + 0.003912920597940683, + 0.00566440261900425, + 0.0041685886681079865, + 0.002261878689751029, + 0.0021808368619531393, + 0.003215139266103506, + 0.004406530410051346, + 0.004546640440821648, + 0.0034272063057869673, + 0.005356760695576668, + 0.0018611181294545531, + 0.001689315540716052, + 0.0024395817890763283, + 0.0030211482662707567, + 0.002751479623839259, + 0.0027890591882169247, + 0.004250458441674709, + 0.003010275075212121, + 0.002263949951156974, + 0.004773429129272699, + 0.004855749197304249, + 0.005667272489517927, + 0.006240934133529663, + 0.009591145440936089, + 0.0050242990255355835, + 0.0072560543194413185, + 0.0032025615219026804, + 0.0038995894137769938, + 0.002406780607998371, + 0.006250984966754913, + 0.003863419173285365, + 0.005045956466346979, + 0.004027387127280235, + 0.004731654655188322, + 0.0023632000666111708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.00528059434145689 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098231438547373 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098231438547373, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098231438547373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3252149820327759, + 0.14277279376983643, + 0.03666115924715996, + 0.29742011427879333, + 0.09005922079086304, + 0.03743599355220795, + 0.018489545211195946, + 0.021029245108366013, + 0.015067175962030888, + 0.015849806368350983 + ], + "credit_map": [ + 0.024390755221247673, + 0.01214007381349802, + 0.11799801141023636, + 0.016705960035324097, + 0.01661469042301178, + 0.1056617721915245, + 0.013725457713007927, + 0.01797826960682869, + 0.0089801624417305, + 0.0047938222996890545, + 0.051245249807834625, + 0.007175028324127197, + 0.009328034706413746, + 0.0424778014421463, + 0.006791991181671619, + 0.011980696581304073, + 0.0042661395855247974, + 0.004460010677576065, + 0.005325388628989458, + 0.004926491528749466, + 0.004367645829916, + 0.004780614748597145, + 0.004157705698162317, + 0.004377161618322134, + 0.017928440123796463, + 0.1098618358373642, + 0.010935289785265923, + 0.0145981190726161, + 0.06006206199526787, + 0.022545255720615387, + 0.01254519633948803, + 0.04894391819834709, + 0.009447803720831871, + 0.012609780766069889, + 0.0219571590423584, + 0.007933308370411396, + 0.02091791294515133, + 0.007591203786432743, + 0.004933149088174105, + 0.0046689012087881565, + 0.0015663134399801493, + 0.007499536033719778, + 0.0033438741229474545, + 0.0038878992199897766, + 0.004981794394552708, + 0.00560796819627285, + 0.005162342451512814, + 0.005386265926063061, + 0.0016683886060491204, + 0.0018274305621162057, + 0.003312824759632349, + 0.003774888813495636, + 0.0019051971612498164, + 0.002183748409152031, + 0.0021436666138470173, + 0.0016734001692384481, + 0.0015566861256957054, + 0.0022146261762827635, + 0.002800570335239172, + 0.002609935821965337, + 0.00281514972448349, + 0.004442322999238968, + 0.0027504200115799904, + 0.001839534263126552, + 0.00174523179885, + 0.0020473019685596228, + 0.001764005864970386, + 0.001880552852526307, + 0.002128810388967395, + 0.0012278604554012418, + 0.0030662601348012686, + 0.0012071524979546666, + 0.0018716122722253203, + 0.001402907189913094, + 0.002419444266706705, + 0.0016560818767175078, + 0.0019703006837517023, + 0.00261083641089499, + 0.0024213234428316355, + 0.0014973008073866367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.004098231438547373 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0205958541482687 + ], + "opposite_logprob_tokens": [ + -3.8955957889556885 + ], + "logprob_reward": -0.0205958541482687, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205958541482687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7718383073806763, + 0.056827571243047714, + 0.013154926709830761, + 0.03313116729259491, + 0.03371600806713104, + 0.031769104301929474, + 0.013681434094905853, + 0.019242476671934128, + 0.019613822922110558, + 0.007025124970823526 + ], + "credit_map": [ + 0.058265045285224915, + 0.022932231426239014, + 0.29812395572662354, + 0.03988567367196083, + 0.031270042061805725, + 0.2472638189792633, + 0.0299519132822752, + 0.044145695865154266, + 0.006624714937061071, + 0.0046072895638644695, + 0.013201854191720486, + 0.004159786272794008, + 0.005849567707628012, + 0.00734084565192461, + 0.005723512265831232, + 0.009320007637143135, + 0.0013872906565666199, + 0.0018242603400722146, + 0.0011009908048436046, + 0.0014442751416936517, + 0.002590714953839779, + 0.0019298410043120384, + 0.001737794023938477, + 0.0011397609487175941, + 0.001981917070224881, + 0.008680740371346474, + 0.0044430638663470745, + 0.0031970094423741102, + 0.0051190610975027084, + 0.0025617999490350485, + 0.0025043687783181667, + 0.004643211606889963, + 0.005093543324619532, + 0.002420877106487751, + 0.010833913460373878, + 0.003013122361153364, + 0.003395749954506755, + 0.0032471828162670135, + 0.0030924833845347166, + 0.0026191403158009052, + 0.0018658267799764872, + 0.0037417516577988863, + 0.0033243445213884115, + 0.004910859279334545, + 0.003673541359603405, + 0.004664605017751455, + 0.005866681225597858, + 0.0037214967887848616, + 0.0013127765851095319, + 0.0012057116255164146, + 0.0024609582033008337, + 0.001886860467493534, + 0.0015634760493412614, + 0.002494731917977333, + 0.001670591882430017, + 0.0010863280622288585, + 0.0013868106761947274, + 0.0020244685001671314, + 0.0027319067157804966, + 0.003414320293813944, + 0.002459512325003743, + 0.0027466791216284037, + 0.002735597314313054, + 0.0017431831220164895, + 0.0022512846626341343, + 0.002353683579713106, + 0.0024632432032376528, + 0.0026040724478662014, + 0.003375000786036253, + 0.0021619931794703007, + 0.003155323676764965, + 0.0012492226669564843, + 0.0006901656161062419, + 0.0005055558867752552, + 0.0010639774845913053, + 0.0010549492435529828, + 0.0012598336907103658, + 0.000850000127684325, + 0.0011221126187592745, + 0.00047853097203187644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0205958541482687 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978105328977108 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.00978105328977108, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978105328977108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1143064945936203, + 0.07415414601564407, + 0.0348268561065197, + 0.2708951532840729, + 0.15645617246627808, + 0.05915510281920433, + 0.04333146661520004, + 0.0704633891582489, + 0.13109469413757324, + 0.045316457748413086 + ], + "credit_map": [ + 0.011844905093312263, + 0.011969768442213535, + 0.02857072651386261, + 0.008388525806367397, + 0.012234690599143505, + 0.0176857803016901, + 0.010919527150690556, + 0.012692577205598354, + 0.005071981810033321, + 0.005610366351902485, + 0.020832182839512825, + 0.00508086709305644, + 0.006650126073509455, + 0.01614898256957531, + 0.0056964438408613205, + 0.009063209407031536, + 0.005045685917139053, + 0.003950582351535559, + 0.004631418734788895, + 0.004593310412019491, + 0.00487274257466197, + 0.00494488887488842, + 0.0032687492202967405, + 0.0035194798838347197, + 0.015973016619682312, + 0.09842535108327866, + 0.012268858030438423, + 0.015151319094002247, + 0.053408559411764145, + 0.019929081201553345, + 0.011223416775465012, + 0.04451558738946915, + 0.013571913354098797, + 0.032493531703948975, + 0.008584585972130299, + 0.013683859258890152, + 0.04662057384848595, + 0.01595139130949974, + 0.014575626701116562, + 0.010974699631333351, + 0.003197323763743043, + 0.007850720547139645, + 0.005581663455814123, + 0.0065498799085617065, + 0.008302256464958191, + 0.008910187520086765, + 0.012220860458910465, + 0.006542219780385494, + 0.004486233927309513, + 0.004946141969412565, + 0.005524956621229649, + 0.004809950478374958, + 0.00635645492002368, + 0.0076072621159255505, + 0.005734441336244345, + 0.0038660308346152306, + 0.005782500375062227, + 0.00742417061701417, + 0.011870184913277626, + 0.009586654603481293, + 0.008323393762111664, + 0.012404493056237698, + 0.00885414332151413, + 0.006217858754098415, + 0.014437601901590824, + 0.015482289716601372, + 0.015379621647298336, + 0.016275059431791306, + 0.026539862155914307, + 0.014710298739373684, + 0.020338894799351692, + 0.007931076921522617, + 0.004747933708131313, + 0.0028872874099761248, + 0.006335400976240635, + 0.006141604855656624, + 0.007344625890254974, + 0.005142472684383392, + 0.00894102267920971, + 0.003776113735511899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.00978105328977108 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins from an extremely low angle, the camera tilting steadily upward to track the billowing black smoke rising violently from the cracked asphalt. Like an ancient beast awakening, this thick, oily column of smoke rises in a chaotic and powerful rhythm, its interior streaked with fiery orange-red sparks that flash through the inky black haze. The edges of the smoke column exhibit a complex, feathery curling structure, twisting and expanding as it spreads toward a gloomy, oppressive gray sky. The ground is littered with indistinct industrial debris and charred debris, hinting at the source of the disaster. The entire scene is devoured by the smoke, with only occasional flashes of fiery orange light emanating from the burning core at the base of the column, providing the sole illumination for this devastating and oppressive scene, creating an atmosphere of apocalyptic grandeur and awe.", + "seed": 26587, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_4/index_981_seed_26587_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.056608053389936686, + "vlm_logprob": -0.056608053389936686, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.056608053389936686, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23267827928066254 + ], + "opposite_logprob_tokens": [ + -1.6076782941818237 + ], + "logprob_reward": -0.23267827928066254, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23267827928066254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1767502725124359, + 0.09812904894351959, + 0.033572934567928314, + 0.047492265701293945, + 0.041663482785224915, + 0.11335626244544983, + 0.049270324409008026, + 0.3541288673877716, + 0.06859008967876434, + 0.01704646646976471 + ], + "credit_map": [ + 0.016360169276595116, + 0.013468481600284576, + 0.030820554122328758, + 0.036865729838609695, + 0.01320113055408001, + 0.015481891110539436, + 0.01944030448794365, + 0.031112005934119225, + 0.010227223858237267, + 0.005622669123113155, + 0.03888317197561264, + 0.008106039837002754, + 0.006113054230809212, + 0.0160535741597414, + 0.005161906126886606, + 0.007961406372487545, + 0.0035901109222322702, + 0.005261371843516827, + 0.0033016011584550142, + 0.0028289894107729197, + 0.005500125233083963, + 0.005604248959571123, + 0.00519083347171545, + 0.0022956521715968847, + 0.005028012674301863, + 0.006285588722676039, + 0.006502124946564436, + 0.004741712473332882, + 0.008444106206297874, + 0.007873691618442535, + 0.0056554717011749744, + 0.002961555263027549, + 0.004236144479364157, + 0.0036896143574267626, + 0.00788105558604002, + 0.003442496992647648, + 0.005860352888703346, + 0.007492691744118929, + 0.005674830172210932, + 0.003386297496035695, + 0.011627004481852055, + 0.01208046730607748, + 0.017393354326486588, + 0.014162487350404263, + 0.013740534894168377, + 0.01752820424735546, + 0.014248169027268887, + 0.01257603894919157, + 0.0037441151216626167, + 0.0030591459944844246, + 0.004916697274893522, + 0.006445608101785183, + 0.008354416117072105, + 0.006347667891532183, + 0.010678930208086967, + 0.005723746493458748, + 0.02882957085967064, + 0.03672002628445625, + 0.03999246656894684, + 0.08770500868558884, + 0.053542543202638626, + 0.03648420050740242, + 0.04583977162837982, + 0.025015274062752724, + 0.010801161639392376, + 0.00504358159378171, + 0.014127870090305805, + 0.006548749282956123, + 0.006116227712482214, + 0.009289434179663658, + 0.008443216793239117, + 0.00821985024958849, + 0.0017247515497729182, + 0.0015463430900126696, + 0.0023378245532512665, + 0.0017971539637073874, + 0.0018841525306925178, + 0.002590337535366416, + 0.002236955100670457, + 0.0029289494268596172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.23267827928066254 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007652727887034416 + ], + "opposite_logprob_tokens": [ + -4.882652759552002 + ], + "logprob_reward": -0.007652727887034416, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007652727887034416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4270920157432556, + 0.1314646154642105, + 0.042454175651073456, + 0.16396592557430267, + 0.055013593286275864, + 0.032385122030973434, + 0.031541332602500916, + 0.07498028874397278, + 0.02135530114173889, + 0.019747626036405563 + ], + "credit_map": [ + 0.023704247549176216, + 0.03931964933872223, + 0.04792969673871994, + 0.09387523680925369, + 0.033845216035842896, + 0.04967663437128067, + 0.04369647055864334, + 0.09504486620426178, + 0.012977360747754574, + 0.006109016481786966, + 0.05710196495056152, + 0.017498089000582695, + 0.007565399631857872, + 0.018215179443359375, + 0.0063320486806333065, + 0.005665559787303209, + 0.004071889445185661, + 0.005301371216773987, + 0.0048974319361150265, + 0.004333973862230778, + 0.0076426551677286625, + 0.005731180310249329, + 0.00708007114008069, + 0.003395603271201253, + 0.01883353665471077, + 0.022088410332798958, + 0.017606401816010475, + 0.01576157845556736, + 0.029854707419872284, + 0.03030102141201496, + 0.020556995645165443, + 0.008963271044194698, + 0.004726475570350885, + 0.0034619320649653673, + 0.0066647822968661785, + 0.007119456306099892, + 0.0077888015657663345, + 0.009650987572968006, + 0.009832716546952724, + 0.005768442992120981, + 0.002793185878545046, + 0.002615636447444558, + 0.0037868621293455362, + 0.004595848266035318, + 0.004271639045327902, + 0.005267226602882147, + 0.005329616833478212, + 0.0037251089233905077, + 0.002680556382983923, + 0.0030592763796448708, + 0.0033867040183395147, + 0.005510293412953615, + 0.004280516877770424, + 0.0055257584899663925, + 0.003360800677910447, + 0.003737429389730096, + 0.006567342672497034, + 0.006258306093513966, + 0.008098263293504715, + 0.01621044985949993, + 0.009870615787804127, + 0.01194523274898529, + 0.00780820706859231, + 0.008221867494285107, + 0.003691550809890032, + 0.0016688700998201966, + 0.004375608637928963, + 0.001988257048651576, + 0.0017051957547664642, + 0.0029580381233245134, + 0.0025460985489189625, + 0.002421681536361575, + 0.0026719216257333755, + 0.0012753871269524097, + 0.0037827936466783285, + 0.002693273825570941, + 0.0021049571223556995, + 0.002427177270874381, + 0.002612550277262926, + 0.0021795639768242836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.007652727887034416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052565219812095165 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.0052565219812095165, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052565219812095165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1129283532500267, + 0.6205494999885559, + 0.05320800468325615, + 0.09974930435419083, + 0.01922507956624031, + 0.02198573388159275, + 0.011628837324678898, + 0.038283247500658035, + 0.012860141694545746, + 0.009581790305674076 + ], + "credit_map": [ + 0.0061528910882771015, + 0.009817291982471943, + 0.01565059833228588, + 0.02151324413716793, + 0.008333241567015648, + 0.01320517435669899, + 0.013145815581083298, + 0.025110097602009773, + 0.04765405133366585, + 0.02076811157166958, + 0.2668624818325043, + 0.14110954105854034, + 0.03458372503519058, + 0.06496058404445648, + 0.02098442055284977, + 0.023626569658517838, + 0.004675770178437233, + 0.005412432365119457, + 0.0063725486397743225, + 0.006763857789337635, + 0.009744163602590561, + 0.008048582822084427, + 0.008715673349797726, + 0.00347497989423573, + 0.012381240725517273, + 0.009426913224160671, + 0.009605776518583298, + 0.008321891538798809, + 0.021479571238160133, + 0.016856424510478973, + 0.015414050780236721, + 0.006263442803174257, + 0.0016319600399583578, + 0.0011321025667712092, + 0.0016650159377604723, + 0.001700578024610877, + 0.003342181211337447, + 0.004918445833027363, + 0.0033003331627696753, + 0.001534464885480702, + 0.0018514110706746578, + 0.0021309670992195606, + 0.0025631627067923546, + 0.0031652157194912434, + 0.0029375830199569464, + 0.0027821867261081934, + 0.003745835740119219, + 0.002809373661875725, + 0.00125638407189399, + 0.0008183238096535206, + 0.0021623466163873672, + 0.002298804232850671, + 0.0014202758902683854, + 0.0015119756571948528, + 0.0013095168396830559, + 0.0008512101485393941, + 0.003257292788475752, + 0.00396454194560647, + 0.004158393945544958, + 0.009762272238731384, + 0.0056512923911213875, + 0.004302735440433025, + 0.004622412845492363, + 0.00256430939771235, + 0.0014514732174575329, + 0.0009485746268182993, + 0.0023216884583234787, + 0.0012434860691428185, + 0.0017733064014464617, + 0.0017091805348172784, + 0.0015424970770254731, + 0.001869935542345047, + 0.0012290964368730783, + 0.0010920028435066342, + 0.0016944363014772534, + 0.0011377554619684815, + 0.0010677457321435213, + 0.0011418663198128343, + 0.001191958668641746, + 0.0010269297054037452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0052565219812095165 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026378821581602097 + ], + "opposite_logprob_tokens": [ + -3.651378870010376 + ], + "logprob_reward": -0.026378821581602097, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026378821581602097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08244000375270844, + 0.24381856620311737, + 0.05766850337386131, + 0.34555861353874207, + 0.04696565493941307, + 0.02249234914779663, + 0.10515934973955154, + 0.04502955079078674, + 0.036296945065259933, + 0.014570482075214386 + ], + "credit_map": [ + 0.010041514411568642, + 0.008286312222480774, + 0.01275600679218769, + 0.010706747882068157, + 0.008149736560881138, + 0.009515471756458282, + 0.007791819982230663, + 0.015192396007478237, + 0.016231535002589226, + 0.013592288829386234, + 0.09043464064598083, + 0.06298533827066422, + 0.01603841222822666, + 0.020654497668147087, + 0.012171730399131775, + 0.011710110120475292, + 0.005740451626479626, + 0.006195634137839079, + 0.008284073323011398, + 0.005965466611087322, + 0.010574298910796642, + 0.007540087681263685, + 0.008121621794998646, + 0.005246875807642937, + 0.04086510092020035, + 0.02051336131989956, + 0.026862135156989098, + 0.030892878770828247, + 0.09174844622612, + 0.07401314377784729, + 0.037385471165180206, + 0.023278091102838516, + 0.004637621808797121, + 0.0027762234676629305, + 0.0043756356462836266, + 0.0064169419929385185, + 0.007180335000157356, + 0.006800755392760038, + 0.010909595526754856, + 0.0038685472682118416, + 0.0022862483747303486, + 0.0024545115884393454, + 0.002423439873382449, + 0.0035015169996768236, + 0.0029456941410899162, + 0.0027483415324240923, + 0.003955944906920195, + 0.002176652429625392, + 0.012218703515827656, + 0.006730814464390278, + 0.022865774109959602, + 0.014287029393017292, + 0.012582196854054928, + 0.011770754121243954, + 0.015169831924140453, + 0.009534236043691635, + 0.003910588566213846, + 0.0042126071639359, + 0.005158470012247562, + 0.005648138467222452, + 0.007467153947800398, + 0.005806280765682459, + 0.007787900045514107, + 0.005038413219153881, + 0.005166814662516117, + 0.003313816385343671, + 0.006747246719896793, + 0.003508583176881075, + 0.004106477368623018, + 0.004470687359571457, + 0.004916383884847164, + 0.00406693248078227, + 0.0016722367145121098, + 0.0012266309931874275, + 0.002972731366753578, + 0.0013671673368662596, + 0.001668594661168754, + 0.0017507694428786635, + 0.0020094069186598063, + 0.0019029435934498906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.026378821581602097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073916219174862 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011073916219174862, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073916219174862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1923932135105133, + 0.34723758697509766, + 0.07440129667520523, + 0.15629027783870697, + 0.05503902956843376, + 0.022589389234781265, + 0.009647996164858341, + 0.06633574515581131, + 0.04354410618543625, + 0.032521288841962814 + ], + "credit_map": [ + 0.01519493106752634, + 0.015516900457441807, + 0.022289736196398735, + 0.03339531272649765, + 0.015004526823759079, + 0.02401401475071907, + 0.02665262296795845, + 0.04032515734434128, + 0.025913232937455177, + 0.013019858859479427, + 0.14285285770893097, + 0.08458993583917618, + 0.019029216840863228, + 0.03615448996424675, + 0.012730902060866356, + 0.012947107665240765, + 0.007144088391214609, + 0.01010886300355196, + 0.007837673649191856, + 0.0074883149936795235, + 0.01348438486456871, + 0.009427492506802082, + 0.013076085597276688, + 0.005834395065903664, + 0.014080685563385487, + 0.017488807439804077, + 0.016222968697547913, + 0.01553145982325077, + 0.023878850042819977, + 0.036387283354997635, + 0.021754629909992218, + 0.010945596732199192, + 0.004069916438311338, + 0.003550396068021655, + 0.005693965591490269, + 0.006684068590402603, + 0.008818749338388443, + 0.01066642813384533, + 0.010625258088111877, + 0.00493024755269289, + 0.0021053655073046684, + 0.0032587323803454638, + 0.00344643322750926, + 0.0044913035817444324, + 0.002344756620004773, + 0.002715666778385639, + 0.002404826460406184, + 0.0018223043298348784, + 0.0009734698105603456, + 0.000656674150377512, + 0.001782955601811409, + 0.0010746680200099945, + 0.0013312463415786624, + 0.0014412978198379278, + 0.0013746768236160278, + 0.0010130072478204966, + 0.005421120207756758, + 0.006787638179957867, + 0.007091496139764786, + 0.016094228252768517, + 0.010149567387998104, + 0.007409023120999336, + 0.008584820665419102, + 0.004797854460775852, + 0.004261315334588289, + 0.004063260741531849, + 0.006268813740462065, + 0.004214098211377859, + 0.004504296462982893, + 0.00680160615593195, + 0.00616637384518981, + 0.00726434076204896, + 0.003661197144538164, + 0.002301090629771352, + 0.00451461598277092, + 0.00437730411067605, + 0.003951826598495245, + 0.004293377976864576, + 0.004994362127035856, + 0.0044275145046412945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.011073916219174862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins from an extremely low angle, the camera tilting steadily upward to track the billowing black smoke rising violently from the cracked asphalt. Like an ancient beast awakening, this thick, oily column of smoke rises in a chaotic and powerful rhythm, its interior streaked with fiery orange-red sparks that flash through the inky black haze. The edges of the smoke column exhibit a complex, feathery curling structure, twisting and expanding as it spreads toward a gloomy, oppressive gray sky. The ground is littered with indistinct industrial debris and charred debris, hinting at the source of the disaster. The entire scene is devoured by the smoke, with only occasional flashes of fiery orange light emanating from the burning core at the base of the column, providing the sole illumination for this devastating and oppressive scene, creating an atmosphere of apocalyptic grandeur and awe.", + "seed": 26588, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_4/index_981_seed_26588_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10106980055570602, + "vlm_logprob": -0.10106980055570602, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.10106980055570602, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4461725056171417 + ], + "opposite_logprob_tokens": [ + -1.0711724758148193 + ], + "logprob_reward": -0.4461725056171417, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4461725056171417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14513055980205536, + 0.07998441159725189, + 0.03746550902724266, + 0.18318088352680206, + 0.04696778953075409, + 0.06152581050992012, + 0.056782763451337814, + 0.02964804880321026, + 0.07546909898519516, + 0.2838451564311981 + ], + "credit_map": [ + 0.011934731155633926, + 0.011915820650756359, + 0.02143648825585842, + 0.015380426310002804, + 0.015018986538052559, + 0.007577682379633188, + 0.01764295995235443, + 0.044223446398973465, + 0.005572037305682898, + 0.0064178635366261005, + 0.023473422974348068, + 0.008087579160928726, + 0.015985017642378807, + 0.00580193055793643, + 0.00725674768909812, + 0.007389804348349571, + 0.002710573375225067, + 0.00368335098028183, + 0.006161883939057589, + 0.00657226424664259, + 0.007119930814951658, + 0.002948212204501033, + 0.0038985535502433777, + 0.004370734095573425, + 0.026229601353406906, + 0.017009614035487175, + 0.048521582037210464, + 0.03319903835654259, + 0.024846311658620834, + 0.009645464830100536, + 0.014570231549441814, + 0.009159023873507977, + 0.009516320191323757, + 0.004473892506211996, + 0.00577390706166625, + 0.004961908794939518, + 0.010842101648449898, + 0.0035943740513175726, + 0.005160513799637556, + 0.0026447672862559557, + 0.005701243877410889, + 0.007527393754571676, + 0.009168627671897411, + 0.009588626213371754, + 0.008631442673504353, + 0.007929438725113869, + 0.0064303032122552395, + 0.006548731122165918, + 0.0048226709477603436, + 0.007037500385195017, + 0.006197280250489712, + 0.010197584517300129, + 0.00813310593366623, + 0.006693351548165083, + 0.009586146101355553, + 0.004115121904760599, + 0.0033664016518741846, + 0.002958639059215784, + 0.005604097153991461, + 0.0032305859494954348, + 0.0041864002123475075, + 0.003465021261945367, + 0.004544529598206282, + 0.0022923708893358707, + 0.013672620989382267, + 0.00603970792144537, + 0.018829070031642914, + 0.0073324828408658504, + 0.007380051072686911, + 0.007989571429789066, + 0.00833616778254509, + 0.005889419466257095, + 0.03358497843146324, + 0.0276569165289402, + 0.04733007028698921, + 0.025860831141471863, + 0.03324856981635094, + 0.03169199824333191, + 0.04325222969055176, + 0.04121954366564751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.4461725056171417 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005977965891361237 + ], + "opposite_logprob_tokens": [ + -5.130978107452393 + ], + "logprob_reward": -0.005977965891361237, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005977965891361237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09668377041816711, + 0.16260495781898499, + 0.31319645047187805, + 0.09179601073265076, + 0.06101720407605171, + 0.028558846563100815, + 0.07841314375400543, + 0.09439589083194733, + 0.016862131655216217, + 0.05647162348031998 + ], + "credit_map": [ + 0.008074590004980564, + 0.007282234728336334, + 0.014418038539588451, + 0.012869007885456085, + 0.011654816567897797, + 0.006106710992753506, + 0.010548136197030544, + 0.025730233639478683, + 0.011420659720897675, + 0.014658449217677116, + 0.04703929275274277, + 0.01715146005153656, + 0.0360153391957283, + 0.00983422715216875, + 0.01435766089707613, + 0.012127876281738281, + 0.021868828684091568, + 0.03466513007879257, + 0.09351788461208344, + 0.029992172494530678, + 0.07305245846509933, + 0.01990870200097561, + 0.018098268657922745, + 0.02209300920367241, + 0.010648266412317753, + 0.005736134946346283, + 0.03132263943552971, + 0.008780819363892078, + 0.01735677942633629, + 0.006279100198298693, + 0.005196251440793276, + 0.006476013921201229, + 0.015629105269908905, + 0.006100377533584833, + 0.0037949029356241226, + 0.0031314105726778507, + 0.01897515170276165, + 0.00558142876252532, + 0.005255627911537886, + 0.0025492017157375813, + 0.0023216966073960066, + 0.003569248830899596, + 0.003884395584464073, + 0.006000115070492029, + 0.004223598632961512, + 0.0033928423654288054, + 0.0030863340944051743, + 0.0020806158427149057, + 0.00769421411678195, + 0.010865036398172379, + 0.011140583083033562, + 0.014903674833476543, + 0.010423575527966022, + 0.008304103277623653, + 0.008932352066040039, + 0.006149602122604847, + 0.009447881951928139, + 0.012598464265465736, + 0.012446529231965542, + 0.012042147107422352, + 0.013436349108815193, + 0.011554733850061893, + 0.013776795007288456, + 0.009092990309000015, + 0.0027888144832104445, + 0.0013781903544440866, + 0.003215125994756818, + 0.0017012672033160925, + 0.0018783633131533861, + 0.001830210560001433, + 0.002431135857477784, + 0.0016390226082876325, + 0.006846655625849962, + 0.004698287229984999, + 0.009963763877749443, + 0.004925750661641359, + 0.007243231404572725, + 0.006178549025207758, + 0.008858736604452133, + 0.007756651379168034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.005977965891361237 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052610281854867935 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.0052610281854867935, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052610281854867935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06422556936740875, + 0.15001551806926727, + 0.20579445362091064, + 0.2141415774822235, + 0.0510716587305069, + 0.07500702887773514, + 0.05489177256822586, + 0.08651191741228104, + 0.02599887177348137, + 0.07234169542789459 + ], + "credit_map": [ + 0.005517432000488043, + 0.007040110416710377, + 0.00846876297146082, + 0.009676146320998669, + 0.008383926935493946, + 0.006812873762100935, + 0.0068483855575323105, + 0.011477934196591377, + 0.011073751375079155, + 0.01054813526570797, + 0.05057879164814949, + 0.013874072581529617, + 0.03406054899096489, + 0.007326070684939623, + 0.012692416086792946, + 0.009861725382506847, + 0.013728728517889977, + 0.01758098043501377, + 0.06511183083057404, + 0.01659073494374752, + 0.05046367645263672, + 0.011488466523587704, + 0.012491416186094284, + 0.018338613212108612, + 0.022747034206986427, + 0.018429046496748924, + 0.049243927001953125, + 0.03337204456329346, + 0.03203459829092026, + 0.016559846699237823, + 0.022275833413004875, + 0.019479231908917427, + 0.009733304381370544, + 0.004629194736480713, + 0.0057810586877167225, + 0.00427074171602726, + 0.011303477920591831, + 0.007299460005015135, + 0.004757022950798273, + 0.003297399030998349, + 0.006599668879061937, + 0.010070701129734516, + 0.009461910463869572, + 0.015489291399717331, + 0.011516682803630829, + 0.008799131959676743, + 0.00789543054997921, + 0.005174211226403713, + 0.0038955812342464924, + 0.004804047755897045, + 0.004518367815762758, + 0.013268337585031986, + 0.00906587578356266, + 0.004017455969005823, + 0.011543992906808853, + 0.0037781153805553913, + 0.00742426747456193, + 0.00983944721519947, + 0.00868295319378376, + 0.010130397975444794, + 0.01263786107301712, + 0.009901485405862331, + 0.016379419714212418, + 0.0115160858258605, + 0.002503588330000639, + 0.0025784822646528482, + 0.003272470086812973, + 0.0032002830412238836, + 0.003587197046726942, + 0.0031187122222036123, + 0.004519256297498941, + 0.0032188824843615294, + 0.006433524191379547, + 0.006562745198607445, + 0.00910186767578125, + 0.006448450963944197, + 0.010581699199974537, + 0.007135313935577869, + 0.015459653921425343, + 0.010618438944220543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.0052610281854867935 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03815456107258797 + ], + "opposite_logprob_tokens": [ + -3.2881546020507812 + ], + "logprob_reward": -0.03815456107258797, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03815456107258797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05315301567316055, + 0.020569829270243645, + 0.6597602367401123, + 0.06275831907987595, + 0.06837012618780136, + 0.03511097654700279, + 0.02935771457850933, + 0.021183865144848824, + 0.016757547855377197, + 0.032978355884552 + ], + "credit_map": [ + 0.00354351126588881, + 0.005022100172936916, + 0.007907532155513763, + 0.0070765032432973385, + 0.006159051787108183, + 0.005635970272123814, + 0.006681280676275492, + 0.011127063073217869, + 0.0015987996011972427, + 0.0021523518953472376, + 0.004083251114934683, + 0.003354760818183422, + 0.0029991413466632366, + 0.001707581803202629, + 0.002013404853641987, + 0.002660536440089345, + 0.0436418242752552, + 0.06518639624118805, + 0.20710580050945282, + 0.06455013900995255, + 0.15652324259281158, + 0.039857931435108185, + 0.039266232401132584, + 0.04362869635224342, + 0.005848223809152842, + 0.006710075307637453, + 0.015397606417536736, + 0.007631376385688782, + 0.01072166208177805, + 0.0059683253057301044, + 0.005039619281888008, + 0.005441427230834961, + 0.015743207186460495, + 0.007213532458990812, + 0.004537543281912804, + 0.005161949433386326, + 0.01998727209866047, + 0.005486288573592901, + 0.006493554916232824, + 0.003746770555153489, + 0.003676904132589698, + 0.004357128404080868, + 0.0069038444198668, + 0.004727463237941265, + 0.004493996500968933, + 0.003946696408092976, + 0.004462684970349073, + 0.0025422556791454554, + 0.0024693708401173353, + 0.00298526999540627, + 0.0034107372630387545, + 0.007025770377367735, + 0.004579724743962288, + 0.0022049476392567158, + 0.004677311517298222, + 0.002004580572247505, + 0.002019875217229128, + 0.002459342358633876, + 0.0030745055992156267, + 0.003243601182475686, + 0.0030368168372660875, + 0.0019660191610455513, + 0.003283378202468157, + 0.0021003263536840677, + 0.002585991518571973, + 0.0014009037986397743, + 0.0034298659302294254, + 0.0018351793987676501, + 0.001701968489214778, + 0.001778462203219533, + 0.0025054272264242172, + 0.0015197494067251682, + 0.0038244668394327164, + 0.003090020501986146, + 0.005455884151160717, + 0.002915774006396532, + 0.0037633925676345825, + 0.0041100322268903255, + 0.005069997161626816, + 0.0047487844713032246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.03815456107258797 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start from an extremely low angle and tilt steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a thick column of black smoke rising violently from cracked asphalt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are fiery orange-red sparks visible streaking through the interior of the smoke column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground littered with indistinct industrial debris and charred debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by occasional flashes of fiery orange light from the base of the smoke column?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097829420119524 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.0097829420119524, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097829420119524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1514686793088913, + 0.0834062322974205, + 0.1403813362121582, + 0.15086780488491058, + 0.06744535267353058, + 0.08886535465717316, + 0.1274421066045761, + 0.07097437977790833, + 0.051340531557798386, + 0.06780819594860077 + ], + "credit_map": [ + 0.010621530935168266, + 0.014096789993345737, + 0.01942114531993866, + 0.018815556541085243, + 0.025111665949225426, + 0.011369037441909313, + 0.018386323004961014, + 0.03364662826061249, + 0.005343001335859299, + 0.0072409301064908504, + 0.024087589234113693, + 0.009481037966907024, + 0.01803022064268589, + 0.0050095790065824986, + 0.0063474467024207115, + 0.007866431958973408, + 0.008768171072006226, + 0.013511314056813717, + 0.046599190682172775, + 0.013332492671906948, + 0.03168385475873947, + 0.008293085731565952, + 0.00788875948637724, + 0.010304465889930725, + 0.018335461616516113, + 0.011917955242097378, + 0.04714677855372429, + 0.01742885448038578, + 0.02432754635810852, + 0.011532275937497616, + 0.010968285612761974, + 0.009210641495883465, + 0.00797204114496708, + 0.005990142002701759, + 0.009332994930446148, + 0.009904900565743446, + 0.012375026009976864, + 0.010899193584918976, + 0.0071049039252102375, + 0.0038661558646708727, + 0.0071213124319911, + 0.00807726476341486, + 0.0146244578063488, + 0.018413158133625984, + 0.012405753135681152, + 0.00849877204746008, + 0.012045133858919144, + 0.007679506205022335, + 0.010297637432813644, + 0.012543701566755772, + 0.012078780680894852, + 0.02747056819498539, + 0.01967417076230049, + 0.011645592749118805, + 0.024670341983437538, + 0.009061316959559917, + 0.009067085571587086, + 0.008705845102667809, + 0.014360846020281315, + 0.006704675033688545, + 0.008281167596578598, + 0.007421629037708044, + 0.008569004945456982, + 0.007864122278988361, + 0.004439713433384895, + 0.005092449486255646, + 0.007148789707571268, + 0.00661068968474865, + 0.007320130709558725, + 0.005850234534591436, + 0.0088071059435606, + 0.0060714189894497395, + 0.008222904987633228, + 0.00609270203858614, + 0.011983595788478851, + 0.007606397848576307, + 0.007530418690294027, + 0.007666892372071743, + 0.009795250371098518, + 0.008910035714507103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0097829420119524 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_5/index_tensor([3483, 3483])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_5/index_tensor([3483, 3483])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..114801f09f409f9e190b569f7cf813198dd3897a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_5/index_tensor([3483, 3483])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An ultra-wide-angle shot, taken from a low, upward angle, with the camera held still, captures the vast, inky black of the night sky, a tranquil speck of tranquility. Countless distant, pinpoint-sized stars shimmer like fine diamond dust, and a faint, milky white band of nebula is faintly visible, adding a sense of depth to the image. A blazing meteor glides majestically and leisurely from one corner of the frame to the other. Caught in high-frame-rate slow motion, the meteor's core appears intensely incandescent, its edges transitioning to molten gold and fiery orange. It trails a magnificent, elongated tail of light, a trail of shimmering particles and incandescent gas that fades from a brilliant gold near the core to deep, ethereal blues and purples as it dissipates into darkness. The movement of the meteor is so slow and clear that every burst of sparks and the graceful dance of every wisp of gas are rendered with crystal clarity, as if time is stretched, immersing the audience in this silent and grand cosmic spectacle. The whole scene is filled with awe-inspiring tranquility and the grandeur of the universe. The strong contrast between light and dark highlights the short but glorious life journey of the meteor.", + "seed": 33084, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_5/index_3483_seed_33084_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.39654689412564037, + "vlm_logprob": -0.39654689412564037, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.39654689412564037, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01928052306175232 + ], + "opposite_logprob_tokens": [ + -4.394280433654785 + ], + "logprob_reward": -0.01928052306175232, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01928052306175232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12264015525579453, + 0.08721383661031723, + 0.11464759707450867, + 0.10395827144384384, + 0.06922180950641632, + 0.14606846868991852, + 0.09139090776443481, + 0.07016001641750336, + 0.0850180983543396, + 0.10968079417943954 + ], + "credit_map": [ + 0.012940343469381332, + 0.008230617269873619, + 0.012509702704846859, + 0.01528273243457079, + 0.01579427160322666, + 0.019482247531414032, + 0.02195025607943535, + 0.016449980437755585, + 0.005030830856412649, + 0.0035975598730146885, + 0.011437934823334217, + 0.010512781329452991, + 0.007864518091082573, + 0.016014622524380684, + 0.019977519288659096, + 0.012778069823980331, + 0.005646213423460722, + 0.008155463263392448, + 0.035416923463344574, + 0.009576781652867794, + 0.008246464654803276, + 0.0302815530449152, + 0.010591389611363411, + 0.00673280144110322, + 0.007998849265277386, + 0.007193275727331638, + 0.018348703160881996, + 0.009694699198007584, + 0.008819538168609142, + 0.02309507317841053, + 0.01313465554267168, + 0.015673479065299034, + 0.005393800791352987, + 0.00780068663880229, + 0.00850480142980814, + 0.008773814886808395, + 0.007090569008141756, + 0.011587308719754219, + 0.007814258337020874, + 0.01225657295435667, + 0.00822562538087368, + 0.00524437241256237, + 0.05361614003777504, + 0.00777365081012249, + 0.004670920316129923, + 0.0502302423119545, + 0.007540397346019745, + 0.00876711867749691, + 0.007804218679666519, + 0.01286669634282589, + 0.015250565484166145, + 0.014901441521942616, + 0.006566651631146669, + 0.01404521707445383, + 0.008607640862464905, + 0.011348470114171505, + 0.006716461386531591, + 0.007413461338728666, + 0.015323483385145664, + 0.008827010169625282, + 0.0034612880554050207, + 0.012205960229039192, + 0.005820949096232653, + 0.010391409508883953, + 0.008830582723021507, + 0.011359316296875477, + 0.013366532512009144, + 0.011559896171092987, + 0.006955647375434637, + 0.010929601266980171, + 0.010494966991245747, + 0.011521558277308941, + 0.008871645666658878, + 0.013663254678249359, + 0.020942341536283493, + 0.015610963106155396, + 0.008686366491019726, + 0.016819801181554794, + 0.009140392765402794, + 0.015946025028824806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.01928052306175232 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.502106785774231 + ], + "opposite_logprob_tokens": [ + -0.25210675597190857 + ], + "logprob_reward": -1.502106785774231, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.502106785774231, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1733042448759079, + 0.06415992230176926, + 0.07589507102966309, + 0.21286650002002716, + 0.07690069079399109, + 0.11755403876304626, + 0.05372891575098038, + 0.05267982929944992, + 0.08178446441888809, + 0.09112636744976044 + ], + "credit_map": [ + 0.012174387462437153, + 0.010388546623289585, + 0.019514210522174835, + 0.023893076926469803, + 0.02189444564282894, + 0.029531054198741913, + 0.03829415515065193, + 0.017614366486668587, + 0.004957681521773338, + 0.004822473973035812, + 0.007940682582557201, + 0.009627081453800201, + 0.00689489021897316, + 0.010188144631683826, + 0.01202014833688736, + 0.007708813529461622, + 0.004157835617661476, + 0.004799336660653353, + 0.016413811594247818, + 0.008675521239638329, + 0.006564540322870016, + 0.01796719804406166, + 0.010124707594513893, + 0.007192112039774656, + 0.011701343581080437, + 0.0150833735242486, + 0.04634306952357292, + 0.019629718735814095, + 0.015315382741391659, + 0.05800021439790726, + 0.03224128112196922, + 0.014552087523043156, + 0.005985765252262354, + 0.006751255597919226, + 0.015138309448957443, + 0.008038608357310295, + 0.00701546436175704, + 0.017298264428973198, + 0.007328864187002182, + 0.009344153106212616, + 0.005811827722936869, + 0.004500747192651033, + 0.042311109602451324, + 0.005904840771108866, + 0.0050943773239851, + 0.041457340121269226, + 0.005291030742228031, + 0.007182754576206207, + 0.003985363524407148, + 0.006759684532880783, + 0.011022691614925861, + 0.006291802041232586, + 0.0037322186399251223, + 0.010731670074164867, + 0.0038581700064241886, + 0.007347310893237591, + 0.004123017191886902, + 0.007323692552745342, + 0.007512474898248911, + 0.007601541467010975, + 0.004993614740669727, + 0.006281543988734484, + 0.0065852883271873, + 0.008258654735982418, + 0.007397259585559368, + 0.012293101288378239, + 0.01279895007610321, + 0.01044043805450201, + 0.006850323639810085, + 0.010739278979599476, + 0.011183190159499645, + 0.010081916116178036, + 0.00836716964840889, + 0.01293266098946333, + 0.0169210247695446, + 0.01269488874822855, + 0.007079336792230606, + 0.01348165050148964, + 0.006318389438092709, + 0.013331230729818344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -1.502106785774231 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199766330420971 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.014199766330420971, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199766330420971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15895462036132812, + 0.12422545999288559, + 0.05947534367442131, + 0.08819591999053955, + 0.06563431024551392, + 0.19371947646141052, + 0.0786484032869339, + 0.0638105496764183, + 0.06606248021125793, + 0.10127337276935577 + ], + "credit_map": [ + 0.010755522176623344, + 0.009443479590117931, + 0.018167905509471893, + 0.017327599227428436, + 0.024198828265070915, + 0.027156852185726166, + 0.0293127354234457, + 0.02259170450270176, + 0.007601309102028608, + 0.005518004763871431, + 0.018557723611593246, + 0.016843244433403015, + 0.010692729614675045, + 0.02794978767633438, + 0.025375686585903168, + 0.011686979793012142, + 0.0038842204958200455, + 0.004785246215760708, + 0.007752265781164169, + 0.007180299609899521, + 0.005673674400895834, + 0.008214018307626247, + 0.014606029726564884, + 0.007379590533673763, + 0.005805076565593481, + 0.006094078533351421, + 0.015726139768958092, + 0.011576909571886063, + 0.006679359823465347, + 0.020700866356492043, + 0.013364993967115879, + 0.008248496800661087, + 0.005656542722135782, + 0.00617604935541749, + 0.01197642832994461, + 0.007934226654469967, + 0.005161463748663664, + 0.013634145259857178, + 0.00729342782869935, + 0.007802024949342012, + 0.011746621690690517, + 0.007257464807480574, + 0.065540611743927, + 0.01347558293491602, + 0.00795349944382906, + 0.061549946665763855, + 0.013369100168347359, + 0.012826642021536827, + 0.005849924404174089, + 0.010258201509714127, + 0.015724411234259605, + 0.010266290977597237, + 0.0051623242907226086, + 0.014604536816477776, + 0.006700405851006508, + 0.010082309134304523, + 0.005409939214587212, + 0.007029815576970577, + 0.010622205212712288, + 0.007713044993579388, + 0.0055545661598443985, + 0.007867641746997833, + 0.011119204573333263, + 0.008494132198393345, + 0.005764109082520008, + 0.011784873902797699, + 0.01211126521229744, + 0.00809309259057045, + 0.004597839433699846, + 0.011677506379783154, + 0.0049574594013392925, + 0.007076333276927471, + 0.0076507218182086945, + 0.015198610723018646, + 0.01839573308825493, + 0.015671169385313988, + 0.007469879928976297, + 0.020394666120409966, + 0.0058732060715556145, + 0.010619386099278927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.014199766330420971 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818695291876793 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.01818695291876793, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818695291876793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20524705946445465, + 0.11817261576652527, + 0.15444864332675934, + 0.1385645568370819, + 0.09702422469854355, + 0.05874861404299736, + 0.06737924367189407, + 0.03809802234172821, + 0.08238624036312103, + 0.0399307981133461 + ], + "credit_map": [ + 0.014909246936440468, + 0.013983144424855709, + 0.030708272010087967, + 0.025005726143717766, + 0.019863346591591835, + 0.04356710612773895, + 0.03506508469581604, + 0.022145124152302742, + 0.009269492700695992, + 0.007631817366927862, + 0.01308965589851141, + 0.018448958173394203, + 0.014098168350756168, + 0.02131473273038864, + 0.0196670088917017, + 0.014652791433036327, + 0.01167677529156208, + 0.014203996397554874, + 0.03867634758353233, + 0.019147641956806183, + 0.01505296304821968, + 0.02093019336462021, + 0.017046095803380013, + 0.017714625224471092, + 0.010609183460474014, + 0.013340357691049576, + 0.020244723185896873, + 0.015284016728401184, + 0.014895463362336159, + 0.024973928928375244, + 0.016875986009836197, + 0.022340897470712662, + 0.006714804098010063, + 0.007102465257048607, + 0.018464816734194756, + 0.011851918883621693, + 0.009371092543005943, + 0.018199849873781204, + 0.011689169332385063, + 0.013630111701786518, + 0.0049828654155135155, + 0.004890508018434048, + 0.010243848897516727, + 0.0061186011880636215, + 0.005445245653390884, + 0.010092484764754772, + 0.007170188240706921, + 0.00980487372726202, + 0.006338912528008223, + 0.007806658744812012, + 0.010315900668501854, + 0.010404026135802269, + 0.005041658878326416, + 0.007898025214672089, + 0.006350508891046047, + 0.013223549351096153, + 0.0033680128399282694, + 0.0036671508569270372, + 0.006453550420701504, + 0.004714443814009428, + 0.003390892408788204, + 0.005655744578689337, + 0.005021834746003151, + 0.005826394539326429, + 0.006621168460696936, + 0.015056577511131763, + 0.010053533129394054, + 0.011460261419415474, + 0.007535765413194895, + 0.008698823861777782, + 0.012328438460826874, + 0.010631672106683254, + 0.004199584014713764, + 0.005538622848689556, + 0.005154728423804045, + 0.006483170669525862, + 0.0035139783285558224, + 0.0047369059175252914, + 0.004146518185734749, + 0.006157290656119585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.01818695291876793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289604425430298 + ], + "opposite_logprob_tokens": [ + -1.0539604425430298 + ], + "logprob_reward": -0.4289604425430298, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289604425430298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20045439898967743, + 0.08015607297420502, + 0.10252392292022705, + 0.057831499725580215, + 0.10587765276432037, + 0.11248357594013214, + 0.07175932079553604, + 0.0663328692317009, + 0.14824311435222626, + 0.0543375238776207 + ], + "credit_map": [ + 0.018710806965827942, + 0.019394682720303535, + 0.027695197612047195, + 0.030500389635562897, + 0.025029798969626427, + 0.031780436635017395, + 0.02817401848733425, + 0.019169054925441742, + 0.0056173414923250675, + 0.006736976560205221, + 0.01225360669195652, + 0.00926926452666521, + 0.009861002676188946, + 0.011290404945611954, + 0.01796305552124977, + 0.007164424285292625, + 0.0048001604154706, + 0.0035639943089336157, + 0.025450171902775764, + 0.009414139203727245, + 0.008434147574007511, + 0.020972417667508125, + 0.02379474975168705, + 0.006094137206673622, + 0.00378150912001729, + 0.002987020183354616, + 0.015070250257849693, + 0.005078193731606007, + 0.003442465327680111, + 0.015181017108261585, + 0.006061287131160498, + 0.006229756399989128, + 0.0074114506132900715, + 0.01198665238916874, + 0.021422408521175385, + 0.0139283062890172, + 0.00820982363075018, + 0.021957671269774437, + 0.009583985432982445, + 0.01137736439704895, + 0.0066457465291023254, + 0.004965917207300663, + 0.039994340389966965, + 0.006042427383363247, + 0.0036930914502590895, + 0.03972138836979866, + 0.003801187500357628, + 0.007619478274136782, + 0.004456338007003069, + 0.008603328838944435, + 0.017113763839006424, + 0.007520243059843779, + 0.004252278711646795, + 0.01651739701628685, + 0.005353085231035948, + 0.007942888885736465, + 0.005329920910298824, + 0.008433154784142971, + 0.011771101504564285, + 0.009539080783724785, + 0.005186012480407953, + 0.01018173061311245, + 0.0063559552654623985, + 0.009535916149616241, + 0.011578653939068317, + 0.030278362333774567, + 0.017388921231031418, + 0.02193913795053959, + 0.011430100537836552, + 0.015274466015398502, + 0.01919526420533657, + 0.021158207207918167, + 0.004470366984605789, + 0.008330565877258778, + 0.009787522256374359, + 0.006217243615537882, + 0.00343637284822762, + 0.011030072346329689, + 0.004084817599505186, + 0.006980560719966888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.4289604425430298 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An ultra-wide-angle shot, taken from a low, upward angle, with the camera held still, captures the vast, inky black of the night sky, a tranquil speck of tranquility. Countless distant, pinpoint-sized stars shimmer like fine diamond dust, and a faint, milky white band of nebula is faintly visible, adding a sense of depth to the image. A blazing meteor glides majestically and leisurely from one corner of the frame to the other. Caught in high-frame-rate slow motion, the meteor's core appears intensely incandescent, its edges transitioning to molten gold and fiery orange. It trails a magnificent, elongated tail of light, a trail of shimmering particles and incandescent gas that fades from a brilliant gold near the core to deep, ethereal blues and purples as it dissipates into darkness. The movement of the meteor is so slow and clear that every burst of sparks and the graceful dance of every wisp of gas are rendered with crystal clarity, as if time is stretched, immersing the audience in this silent and grand cosmic spectacle. The whole scene is filled with awe-inspiring tranquility and the grandeur of the universe. The strong contrast between light and dark highlights the short but glorious life journey of the meteor.", + "seed": 33085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_5/index_3483_seed_33085_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.30471874475479127, + "vlm_logprob": -0.30471874475479127, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.30471874475479127, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03206770122051239 + ], + "opposite_logprob_tokens": [ + -4.157067775726318 + ], + "logprob_reward": -0.03206770122051239, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03206770122051239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12349291890859604, + 0.046166058629751205, + 0.06564366072416306, + 0.025067152455449104, + 0.382284939289093, + 0.07278121262788773, + 0.08592779189348221, + 0.08946506679058075, + 0.06575054675340652, + 0.04342065006494522 + ], + "credit_map": [ + 0.009815050289034843, + 0.006863920483738184, + 0.017912114039063454, + 0.011019271798431873, + 0.0276875589042902, + 0.025623813271522522, + 0.008895913138985634, + 0.015675265341997147, + 0.0038574952632188797, + 0.0026310537941753864, + 0.0048818038776516914, + 0.005614478141069412, + 0.009641329757869244, + 0.010827356949448586, + 0.003953676670789719, + 0.004758861381560564, + 0.003796083852648735, + 0.002674480900168419, + 0.008677373640239239, + 0.006582129281014204, + 0.01512082852423191, + 0.01585194282233715, + 0.007207984104752541, + 0.005732837598770857, + 0.0017821458168327808, + 0.002306334674358368, + 0.005151879508048296, + 0.003069869475439191, + 0.0028534538578242064, + 0.0033600409515202045, + 0.002166120335459709, + 0.004377306438982487, + 0.011836017481982708, + 0.005884675309062004, + 0.13860374689102173, + 0.01519738882780075, + 0.031573716551065445, + 0.16042324900627136, + 0.00540247093886137, + 0.013363637030124664, + 0.00486509595066309, + 0.012211737222969532, + 0.010842658579349518, + 0.00916121806949377, + 0.008167371153831482, + 0.009864754974842072, + 0.006518527399748564, + 0.011149846017360687, + 0.005483833607286215, + 0.018531104549765587, + 0.011549771763384342, + 0.011277749203145504, + 0.008636834099888802, + 0.01813805289566517, + 0.00519583048298955, + 0.007114614360034466, + 0.006114152260124683, + 0.018445611000061035, + 0.00802566483616829, + 0.009329739958047867, + 0.014560241252183914, + 0.014889969490468502, + 0.009010265581309795, + 0.00908941775560379, + 0.004286481998860836, + 0.015015331096947193, + 0.00802921038120985, + 0.003364147152751684, + 0.01631716452538967, + 0.011229484342038631, + 0.0031743375584483147, + 0.004334379453212023, + 0.0022026903461664915, + 0.01619011163711548, + 0.0024000138510018587, + 0.0018865662859752774, + 0.012628498487174511, + 0.0034475161228328943, + 0.0017236623680219054, + 0.0029415825847536325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.03206770122051239 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.898345410823822 + ], + "opposite_logprob_tokens": [ + -0.523345410823822 + ], + "logprob_reward": -0.898345410823822, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.898345410823822, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.080936960875988, + 0.0699763298034668, + 0.03196616843342781, + 0.043324064463377, + 0.13799498975276947, + 0.12118726968765259, + 0.08733518421649933, + 0.31588292121887207, + 0.09541372209787369, + 0.015982307493686676 + ], + "credit_map": [ + 0.009343216195702553, + 0.00441389624029398, + 0.010396326892077923, + 0.007293423637747765, + 0.015146530233323574, + 0.015013647265732288, + 0.00539135467261076, + 0.013938570395112038, + 0.005775984842330217, + 0.004105076193809509, + 0.01288676168769598, + 0.009894200600683689, + 0.01074085384607315, + 0.011816093698143959, + 0.006373489275574684, + 0.008383865468204021, + 0.0024331368040293455, + 0.002045737113803625, + 0.004270536359399557, + 0.004207806661725044, + 0.005645486526191235, + 0.005561825819313526, + 0.003355613211169839, + 0.004446027334779501, + 0.0038968774024397135, + 0.0034812826197594404, + 0.005125442054122686, + 0.004319838713854551, + 0.006924473214894533, + 0.011078675277531147, + 0.004401087295264006, + 0.004096388351172209, + 0.0042999121360480785, + 0.002747457940131426, + 0.048450734466314316, + 0.005319129675626755, + 0.012074277736246586, + 0.055818866938352585, + 0.0041327327489852905, + 0.005151880439370871, + 0.006591008976101875, + 0.01509213075041771, + 0.016188886016607285, + 0.025375178083777428, + 0.016672974452376366, + 0.015339423902332783, + 0.010848081670701504, + 0.015079597011208534, + 0.005519084166735411, + 0.015361656434834003, + 0.018155137076973915, + 0.008545721881091595, + 0.007996346801519394, + 0.020052039995789528, + 0.005177854094654322, + 0.0065273381769657135, + 0.018810709938406944, + 0.09239791333675385, + 0.03455503657460213, + 0.025142857804894447, + 0.04486265406012535, + 0.0512111559510231, + 0.024750974029302597, + 0.024151621386408806, + 0.007238697726279497, + 0.012373105622828007, + 0.010667002759873867, + 0.012814969755709171, + 0.014164038002490997, + 0.019594207406044006, + 0.008174716494977474, + 0.010386982001364231, + 0.0010388904483988881, + 0.0036929259076714516, + 0.0014184090541675687, + 0.0014386688126251101, + 0.0036991359665989876, + 0.0020365144591778517, + 0.0009928782237693667, + 0.001664882991462946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.898345410823822 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953080952167511 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953080952167511, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953080952167511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05523901805281639, + 0.05757253244519234, + 0.06139694154262543, + 0.052979715168476105, + 0.2496354728937149, + 0.18135540187358856, + 0.09609820693731308, + 0.17931054532527924, + 0.049044474959373474, + 0.01736774668097496 + ], + "credit_map": [ + 0.00473440857604146, + 0.002888674149289727, + 0.006508059799671173, + 0.005102385301142931, + 0.012965458445250988, + 0.010793443769216537, + 0.004855130333453417, + 0.00739145465195179, + 0.004987796302884817, + 0.0045397221110761166, + 0.008684785105288029, + 0.006826183758676052, + 0.009433107450604439, + 0.011429271660745144, + 0.00568107794970274, + 0.0059905871748924255, + 0.0047887172549963, + 0.0038371235132217407, + 0.009032730013132095, + 0.007518162950873375, + 0.010888342745602131, + 0.009895838797092438, + 0.007054287474602461, + 0.008381733670830727, + 0.004284766037017107, + 0.005627598147839308, + 0.008159039542078972, + 0.006389327347278595, + 0.007076126988977194, + 0.0105091268196702, + 0.00597034627571702, + 0.004963374696671963, + 0.00762347225099802, + 0.005887526087462902, + 0.09028705954551697, + 0.011104639619588852, + 0.019832022488117218, + 0.10214414447546005, + 0.004418623633682728, + 0.008337968029081821, + 0.012016433291137218, + 0.029230671003460884, + 0.029764113947749138, + 0.031022071838378906, + 0.020712750032544136, + 0.020889898762106895, + 0.015479939058423042, + 0.022239521145820618, + 0.00488917576149106, + 0.023438189178705215, + 0.015017680823802948, + 0.010505368001759052, + 0.012003524228930473, + 0.014416203834116459, + 0.00879671424627304, + 0.007031353190541267, + 0.0084330840036273, + 0.05968429148197174, + 0.042908914387226105, + 0.012132275849580765, + 0.015440921299159527, + 0.020829463377594948, + 0.009038231335580349, + 0.010843338444828987, + 0.0027349325828254223, + 0.007993504405021667, + 0.006302672903984785, + 0.005072332452982664, + 0.012587233446538448, + 0.0066888658329844475, + 0.003420652588829398, + 0.004244272597134113, + 0.0007466102251783013, + 0.0038950678426772356, + 0.0014477476943284273, + 0.0017764541553333402, + 0.0036520566791296005, + 0.0032017177436500788, + 0.0011640740558505058, + 0.0014840175863355398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.005953080952167511 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292148381471634 + ], + "opposite_logprob_tokens": [ + -2.2379214763641357 + ], + "logprob_reward": -0.11292148381471634, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292148381471634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15325170755386353, + 0.08960853517055511, + 0.06937707960605621, + 0.05199623107910156, + 0.2201521396636963, + 0.1356879025697708, + 0.08228377997875214, + 0.046578578650951385, + 0.12272737175226212, + 0.028336644172668457 + ], + "credit_map": [ + 0.016569072380661964, + 0.009870903566479683, + 0.020464610308408737, + 0.014509587548673153, + 0.030518120154738426, + 0.029690681025385857, + 0.011293373070657253, + 0.020335359498858452, + 0.0065398490987718105, + 0.004232152830809355, + 0.009873508475720882, + 0.007830710150301456, + 0.022099152207374573, + 0.023610573261976242, + 0.00689686369150877, + 0.008525723591446877, + 0.005341243464499712, + 0.0029064177069813013, + 0.012240789830684662, + 0.007462461479008198, + 0.012045580893754959, + 0.014234457165002823, + 0.006680179387331009, + 0.008465946651995182, + 0.005057922564446926, + 0.006654706783592701, + 0.008574333041906357, + 0.007152410689741373, + 0.00628335727378726, + 0.00825427658855915, + 0.0046305046416819096, + 0.005388720892369747, + 0.006707891821861267, + 0.004042959772050381, + 0.07778133451938629, + 0.008550143800675869, + 0.01934823766350746, + 0.09171734750270844, + 0.00502379285171628, + 0.006980425678193569, + 0.007284000515937805, + 0.018654990941286087, + 0.01984185166656971, + 0.024921592324972153, + 0.01718808338046074, + 0.019892171025276184, + 0.01189781166613102, + 0.016007399186491966, + 0.00499351741746068, + 0.012868699617683887, + 0.016205724328756332, + 0.00823436863720417, + 0.009388040751218796, + 0.01728452369570732, + 0.006709318608045578, + 0.006599589250981808, + 0.002552604768425226, + 0.006142236292362213, + 0.004621948581188917, + 0.004589409567415714, + 0.008676146157085896, + 0.00884280726313591, + 0.005503466352820396, + 0.005649959202855825, + 0.009767199866473675, + 0.01964454911649227, + 0.01300954818725586, + 0.014262263663113117, + 0.018202444538474083, + 0.026183195412158966, + 0.00990517158061266, + 0.011752991005778313, + 0.0020885937847197056, + 0.0031687943264842033, + 0.00334050296805799, + 0.0036885151639580727, + 0.003616931615397334, + 0.005361214745789766, + 0.003056362271308899, + 0.004015730228275061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.11292148381471634 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47430604696273804 + ], + "opposite_logprob_tokens": [ + -0.974306046962738 + ], + "logprob_reward": -0.47430604696273804, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47430604696273804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13088086247444153, + 0.08017793297767639, + 0.02781904675066471, + 0.024629788473248482, + 0.04832037165760994, + 0.05093901604413986, + 0.1244475468993187, + 0.32129332423210144, + 0.16960106790065765, + 0.021891014650464058 + ], + "credit_map": [ + 0.01129903458058834, + 0.005409421864897013, + 0.014357156120240688, + 0.009684573858976364, + 0.030868196859955788, + 0.030883371829986572, + 0.010118626058101654, + 0.018260477110743523, + 0.00806411448866129, + 0.0036491260398179293, + 0.010666899383068085, + 0.007166904862970114, + 0.01550734881311655, + 0.018496425822377205, + 0.005178945139050484, + 0.01144817192107439, + 0.0020283509511500597, + 0.0017599015263840556, + 0.00398155627772212, + 0.003463044995442033, + 0.004441099241375923, + 0.00446972856298089, + 0.0035997761879116297, + 0.004075587261468172, + 0.002350067486986518, + 0.002170094521716237, + 0.003190655494108796, + 0.002345760352909565, + 0.003143167123198509, + 0.0037903243210166693, + 0.0024177285376936197, + 0.005221990868449211, + 0.0025387441273778677, + 0.004038674756884575, + 0.011034451425075531, + 0.0038794309366494417, + 0.004774945322424173, + 0.013608119450509548, + 0.003431027987971902, + 0.005014976020902395, + 0.003748595714569092, + 0.008551829494535923, + 0.009788860566914082, + 0.006974499672651291, + 0.004772141110152006, + 0.008746231906116009, + 0.0037099523469805717, + 0.004646908491849899, + 0.0077258264645934105, + 0.023950859904289246, + 0.020384345203638077, + 0.014077059924602509, + 0.012126515619456768, + 0.023421915248036385, + 0.009914875961840153, + 0.012846163474023342, + 0.01768324337899685, + 0.10904272645711899, + 0.04327087849378586, + 0.02274218015372753, + 0.03851073980331421, + 0.04713431000709534, + 0.021461479365825653, + 0.021447764709591866, + 0.011826178058981895, + 0.04098406434059143, + 0.01429283432662487, + 0.01412111334502697, + 0.035760413855314255, + 0.022999370470643044, + 0.009289398789405823, + 0.020327696576714516, + 0.0016719213454052806, + 0.004074112046509981, + 0.0025192322209477425, + 0.0027634305879473686, + 0.0035935011692345142, + 0.003473718650639057, + 0.0018825476290658116, + 0.0019125508842989802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.47430604696273804 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An ultra-wide-angle shot, taken from a low, upward angle, with the camera held still, captures the vast, inky black of the night sky, a tranquil speck of tranquility. Countless distant, pinpoint-sized stars shimmer like fine diamond dust, and a faint, milky white band of nebula is faintly visible, adding a sense of depth to the image. A blazing meteor glides majestically and leisurely from one corner of the frame to the other. Caught in high-frame-rate slow motion, the meteor's core appears intensely incandescent, its edges transitioning to molten gold and fiery orange. It trails a magnificent, elongated tail of light, a trail of shimmering particles and incandescent gas that fades from a brilliant gold near the core to deep, ethereal blues and purples as it dissipates into darkness. The movement of the meteor is so slow and clear that every burst of sparks and the graceful dance of every wisp of gas are rendered with crystal clarity, as if time is stretched, immersing the audience in this silent and grand cosmic spectacle. The whole scene is filled with awe-inspiring tranquility and the grandeur of the universe. The strong contrast between light and dark highlights the short but glorious life journey of the meteor.", + "seed": 33086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_5/index_3483_seed_33086_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3483028536662459, + "vlm_logprob": -0.3483028536662459, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.3483028536662459, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02016870118677616 + ], + "opposite_logprob_tokens": [ + -4.395168781280518 + ], + "logprob_reward": -0.02016870118677616, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02016870118677616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16531209647655487, + 0.06765832751989365, + 0.15821120142936707, + 0.06440606713294983, + 0.0450778491795063, + 0.18579673767089844, + 0.09903837740421295, + 0.09218832850456238, + 0.030185598880052567, + 0.09212543815374374 + ], + "credit_map": [ + 0.010055185295641422, + 0.01570923626422882, + 0.01999189332127571, + 0.023993059992790222, + 0.02726423367857933, + 0.026631828397512436, + 0.021499406546354294, + 0.020167235285043716, + 0.005204963963478804, + 0.005846499465405941, + 0.0126102976500988, + 0.006412104703485966, + 0.009698612615466118, + 0.009460389614105225, + 0.007892546243965626, + 0.01053291093558073, + 0.008982167579233646, + 0.010325624607503414, + 0.019407445564866066, + 0.009920304641127586, + 0.03211745247244835, + 0.045992858707904816, + 0.01329930778592825, + 0.018166042864322662, + 0.004215552471578121, + 0.0031581458169966936, + 0.010756351053714752, + 0.005656872410327196, + 0.004861556924879551, + 0.010150275193154812, + 0.016920924186706543, + 0.008686385117471218, + 0.0047737048007547855, + 0.0031588044948875904, + 0.009411453269422054, + 0.005688635632395744, + 0.0036205213982611895, + 0.0068926140666007996, + 0.005304909311234951, + 0.006227202247828245, + 0.014666328206658363, + 0.04761861264705658, + 0.02339955046772957, + 0.034642335027456284, + 0.008786004036664963, + 0.01666688546538353, + 0.022533979266881943, + 0.017483048141002655, + 0.009596417658030987, + 0.012005879543721676, + 0.01528832409530878, + 0.01873713731765747, + 0.005836134776473045, + 0.010172558017075062, + 0.0177188478410244, + 0.009683066047728062, + 0.009725871495902538, + 0.011601139791309834, + 0.013538522645831108, + 0.011272850446403027, + 0.012159980833530426, + 0.010369391180574894, + 0.015186883509159088, + 0.008333680219948292, + 0.0030670438427478075, + 0.003161852713674307, + 0.003845326602458954, + 0.004496757872402668, + 0.003311443142592907, + 0.003718577791005373, + 0.004842957016080618, + 0.003741637570783496, + 0.0060435570776462555, + 0.008887928910553455, + 0.009460972622036934, + 0.017725801095366478, + 0.020983902737498283, + 0.008802625350654125, + 0.009844908490777016, + 0.010375738143920898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.02016870118677616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063427448272705 + ], + "opposite_logprob_tokens": [ + -0.28134268522262573 + ], + "logprob_reward": -1.4063427448272705, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063427448272705, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16020186245441437, + 0.07360328733921051, + 0.21359896659851074, + 0.06927222013473511, + 0.10571206361055374, + 0.16911283135414124, + 0.035545893013477325, + 0.06472260504961014, + 0.061178624629974365, + 0.04705171659588814 + ], + "credit_map": [ + 0.011801505461335182, + 0.009868593886494637, + 0.01954573206603527, + 0.010990286245942116, + 0.0474432073533535, + 0.03375423327088356, + 0.013230731710791588, + 0.01356755942106247, + 0.005927678197622299, + 0.01126853283494711, + 0.011866305023431778, + 0.008105415850877762, + 0.011690560728311539, + 0.00987932551652193, + 0.008044864982366562, + 0.0068206037394702435, + 0.01919207163155079, + 0.02326226606965065, + 0.028205247595906258, + 0.018238615244627, + 0.031147832050919533, + 0.03020446188747883, + 0.0336194783449173, + 0.02972896210849285, + 0.006075317971408367, + 0.004058379214257002, + 0.011791455559432507, + 0.007757894229143858, + 0.009715552441775799, + 0.009020811878144741, + 0.01345392968505621, + 0.007398870773613453, + 0.010918947868049145, + 0.008532541804015636, + 0.02079872041940689, + 0.007264862302690744, + 0.006062845233827829, + 0.027489816769957542, + 0.009260035119950771, + 0.015384281985461712, + 0.01348742563277483, + 0.04206134006381035, + 0.019435066729784012, + 0.028243592008948326, + 0.0070038107223808765, + 0.017917783930897713, + 0.022454360499978065, + 0.018509427085518837, + 0.00354185258038342, + 0.004442387726157904, + 0.004426207393407822, + 0.004733724519610405, + 0.002846447052434087, + 0.004671139642596245, + 0.007026721723377705, + 0.003857410280033946, + 0.006571909878402948, + 0.008609757758677006, + 0.00836936105042696, + 0.012571975588798523, + 0.007365517783910036, + 0.0051400018855929375, + 0.010136796161532402, + 0.005957282148301601, + 0.006929274182766676, + 0.007926535792648792, + 0.008110121823847294, + 0.00790217611938715, + 0.010186935774981976, + 0.006025637034326792, + 0.008622616529464722, + 0.005475323647260666, + 0.004917006008327007, + 0.007012689486145973, + 0.0070639559999108315, + 0.006830563303083181, + 0.008141035214066505, + 0.004362063482403755, + 0.004565555136650801, + 0.004158845636993647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -1.4063427448272705 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208228327333927 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.014208228327333927, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208228327333927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1140923872590065, + 0.10377994179725647, + 0.08071248978376389, + 0.06025309860706329, + 0.12101572006940842, + 0.2205132097005844, + 0.13328143954277039, + 0.11275366693735123, + 0.034120552241802216, + 0.019477443769574165 + ], + "credit_map": [ + 0.008532523177564144, + 0.01087335217744112, + 0.012987644411623478, + 0.014645136892795563, + 0.020116625353693962, + 0.017505832016468048, + 0.015579547733068466, + 0.013851728290319443, + 0.009177180007100105, + 0.014075200073421001, + 0.017520667985081673, + 0.01070861704647541, + 0.01457437314093113, + 0.012851743958890438, + 0.014829163439571857, + 0.010042991489171982, + 0.005632629618048668, + 0.004315240308642387, + 0.010547060519456863, + 0.004347777459770441, + 0.014997805468738079, + 0.02034139819443226, + 0.010400790721178055, + 0.01012978982180357, + 0.005379895213991404, + 0.003495641518384218, + 0.00905741099268198, + 0.005603565834462643, + 0.00932038202881813, + 0.008678861893713474, + 0.0117618003860116, + 0.006955541670322418, + 0.010607418604195118, + 0.009055456146597862, + 0.02599645033478737, + 0.013456013984978199, + 0.008955923840403557, + 0.029794570058584213, + 0.013158546760678291, + 0.009991341270506382, + 0.019150424748659134, + 0.06258989870548248, + 0.029670942574739456, + 0.026613391935825348, + 0.01423205342143774, + 0.02215190976858139, + 0.02281036786735058, + 0.023294223472476006, + 0.012243514880537987, + 0.015688428655266762, + 0.02033437415957451, + 0.020804516971111298, + 0.010340407490730286, + 0.015889698639512062, + 0.023581359535455704, + 0.014399142935872078, + 0.0118261082097888, + 0.01764684170484543, + 0.015325887128710747, + 0.016756586730480194, + 0.012061873450875282, + 0.010422615334391594, + 0.018383987247943878, + 0.010329770855605602, + 0.0035481739323586226, + 0.004419750068336725, + 0.004118325188755989, + 0.0047400048933923244, + 0.003956439904868603, + 0.004365530796349049, + 0.0059985993430018425, + 0.002973729744553566, + 0.0016467327950522304, + 0.002286833943799138, + 0.0025039827451109886, + 0.0034112269058823586, + 0.0029721944592893124, + 0.002192472340539098, + 0.002497113076969981, + 0.001966887852177024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.014208228327333927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865875095129013 + ], + "opposite_logprob_tokens": [ + -3.048658847808838 + ], + "logprob_reward": -0.04865875095129013, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865875095129013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1650804877281189, + 0.15767578780651093, + 0.06308165937662125, + 0.02411697618663311, + 0.07423504441976547, + 0.07968666404485703, + 0.05233479663729668, + 0.06969581544399261, + 0.16039276123046875, + 0.15370012819766998 + ], + "credit_map": [ + 0.0185354333370924, + 0.023488910868763924, + 0.023916592821478844, + 0.014305328018963337, + 0.02498021349310875, + 0.024006150662899017, + 0.012841148301959038, + 0.02300667203962803, + 0.010405016131699085, + 0.01353830099105835, + 0.030185505747795105, + 0.010905102826654911, + 0.030119983479380608, + 0.02280167117714882, + 0.022038618102669716, + 0.017681561410427094, + 0.005380637012422085, + 0.006793236825615168, + 0.010478745214641094, + 0.00535788107663393, + 0.008477091789245605, + 0.0110098198056221, + 0.008128703571856022, + 0.007455532904714346, + 0.0021749590523540974, + 0.001734927180223167, + 0.0039634136483073235, + 0.0032688488718122244, + 0.0021068817004561424, + 0.004274036269634962, + 0.0031434972770512104, + 0.003450407413765788, + 0.010093427263200283, + 0.009423332288861275, + 0.010989892296493053, + 0.008668337017297745, + 0.006872243247926235, + 0.0091102235019207, + 0.009082062169909477, + 0.00999551173299551, + 0.0075560263358056545, + 0.020779749378561974, + 0.011816281825304031, + 0.009895376861095428, + 0.006011905614286661, + 0.007391834631562233, + 0.008286225609481335, + 0.007949254475533962, + 0.005262509919703007, + 0.007650257553905249, + 0.008905911818146706, + 0.007308077532798052, + 0.0042278701439499855, + 0.005662649869918823, + 0.007576948497444391, + 0.005740560125559568, + 0.008538784459233284, + 0.010168815962970257, + 0.008457318879663944, + 0.008505997247993946, + 0.010448777116835117, + 0.007171613164246082, + 0.010439850389957428, + 0.005964649375528097, + 0.016782382503151894, + 0.02026374265551567, + 0.020024504512548447, + 0.019224926829338074, + 0.02557036094367504, + 0.021394921466708183, + 0.021079787984490395, + 0.016052110120654106, + 0.008971987292170525, + 0.01562749594449997, + 0.017401091754436493, + 0.02600262127816677, + 0.03737052157521248, + 0.015105840750038624, + 0.01746419072151184, + 0.015756353735923767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.04865875095129013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25213584303855896 + ], + "opposite_logprob_tokens": [ + -1.5021358728408813 + ], + "logprob_reward": -0.25213584303855896, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25213584303855896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21325592696666718, + 0.1444820761680603, + 0.10540449619293213, + 0.05946328863501549, + 0.041023559868335724, + 0.06702254712581635, + 0.03138264641165733, + 0.1502637416124344, + 0.12524528801441193, + 0.06245643272995949 + ], + "credit_map": [ + 0.013974354602396488, + 0.018445800989866257, + 0.02815103903412819, + 0.028860153630375862, + 0.04189879074692726, + 0.04064363241195679, + 0.022342000156641006, + 0.018940145149827003, + 0.009273845702409744, + 0.014132591895759106, + 0.013470280915498734, + 0.009073206223547459, + 0.04390599578619003, + 0.03624403104186058, + 0.010033229365944862, + 0.00834888219833374, + 0.0076779648661613464, + 0.010977940633893013, + 0.020208068192005157, + 0.008843880146741867, + 0.01264080312103033, + 0.017225533723831177, + 0.016818027943372726, + 0.011012277565896511, + 0.004602246452122927, + 0.0027486789040267467, + 0.005113955121487379, + 0.0048477258533239365, + 0.005604211241006851, + 0.008292460814118385, + 0.017947431653738022, + 0.010306581854820251, + 0.0041143023408949375, + 0.002993115922436118, + 0.007182423956692219, + 0.003449451643973589, + 0.0033641939517110586, + 0.007003309205174446, + 0.005444471258670092, + 0.007472289726138115, + 0.006373590789735317, + 0.007985630072653294, + 0.010559512302279472, + 0.011348885484039783, + 0.004684080369770527, + 0.0075917295180261135, + 0.010550710372626781, + 0.007928409613668919, + 0.002811112441122532, + 0.002878003753721714, + 0.004637342877686024, + 0.006361930165439844, + 0.0030947430059313774, + 0.0032874862663447857, + 0.005294626113027334, + 0.0030174010898917913, + 0.01677098497748375, + 0.01972940005362034, + 0.020379068329930305, + 0.030276689678430557, + 0.02141513302922249, + 0.011758037842810154, + 0.020646164193749428, + 0.009288258850574493, + 0.013561361469328403, + 0.016706349328160286, + 0.016201090067625046, + 0.018098561093211174, + 0.016083890572190285, + 0.016612714156508446, + 0.015725575387477875, + 0.012255745939910412, + 0.00399342505261302, + 0.00615641288459301, + 0.00761374132707715, + 0.01158448401838541, + 0.014209367334842682, + 0.005615490023046732, + 0.006749818101525307, + 0.006533689331263304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.25213584303855896 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An ultra-wide-angle shot, taken from a low, upward angle, with the camera held still, captures the vast, inky black of the night sky, a tranquil speck of tranquility. Countless distant, pinpoint-sized stars shimmer like fine diamond dust, and a faint, milky white band of nebula is faintly visible, adding a sense of depth to the image. A blazing meteor glides majestically and leisurely from one corner of the frame to the other. Caught in high-frame-rate slow motion, the meteor's core appears intensely incandescent, its edges transitioning to molten gold and fiery orange. It trails a magnificent, elongated tail of light, a trail of shimmering particles and incandescent gas that fades from a brilliant gold near the core to deep, ethereal blues and purples as it dissipates into darkness. The movement of the meteor is so slow and clear that every burst of sparks and the graceful dance of every wisp of gas are rendered with crystal clarity, as if time is stretched, immersing the audience in this silent and grand cosmic spectacle. The whole scene is filled with awe-inspiring tranquility and the grandeur of the universe. The strong contrast between light and dark highlights the short but glorious life journey of the meteor.", + "seed": 33087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_5/index_3483_seed_33087_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.35440573543310167, + "vlm_logprob": -0.35440573543310167, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.35440573543310167, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03301645442843437 + ], + "opposite_logprob_tokens": [ + -3.7830164432525635 + ], + "logprob_reward": -0.03301645442843437, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03301645442843437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08076272159814835, + 0.03386745601892471, + 0.03901186212897301, + 0.39162150025367737, + 0.031184036284685135, + 0.19141095876693726, + 0.05930906534194946, + 0.0769214779138565, + 0.02018226869404316, + 0.0757286325097084 + ], + "credit_map": [ + 0.01266423985362053, + 0.0068201529793441296, + 0.01100757997483015, + 0.006228167098015547, + 0.0065392362885177135, + 0.01260993629693985, + 0.00793664064258337, + 0.01695677451789379, + 0.004100707825273275, + 0.001863145618699491, + 0.006375311408191919, + 0.007053757086396217, + 0.0014472315087914467, + 0.005455817561596632, + 0.003710303921252489, + 0.0038611835334450006, + 0.0038859283085912466, + 0.003176555735990405, + 0.004877025261521339, + 0.00448540598154068, + 0.0026663977187126875, + 0.009396011009812355, + 0.0050378115847706795, + 0.005486730951815844, + 0.021912993863224983, + 0.13036686182022095, + 0.1483515053987503, + 0.02797968126833439, + 0.012558754533529282, + 0.029367469251155853, + 0.009202392771840096, + 0.011881863698363304, + 0.002596569713205099, + 0.005085007753223181, + 0.004364789463579655, + 0.004711925517767668, + 0.0031301744747906923, + 0.0032241367734968662, + 0.00532137043774128, + 0.0027500654105097055, + 0.013650826178491116, + 0.04937581717967987, + 0.0781838595867157, + 0.01524722296744585, + 0.008025603368878365, + 0.01507343165576458, + 0.005645443219691515, + 0.006208751816302538, + 0.0038556328509002924, + 0.01498901005834341, + 0.018330764025449753, + 0.0057656788267195225, + 0.003932516556233168, + 0.006061895750463009, + 0.0031448877416551113, + 0.003228682791814208, + 0.004095584154129028, + 0.019675610587000847, + 0.02386942133307457, + 0.0056901536881923676, + 0.0053142667748034, + 0.007841438055038452, + 0.0038148046005517244, + 0.00662019569426775, + 0.0013195928186178207, + 0.0039850384928286076, + 0.005112296435981989, + 0.002381098922342062, + 0.001797245698980987, + 0.002183946082368493, + 0.0021705529652535915, + 0.0012324987910687923, + 0.005542170722037554, + 0.02040952257812023, + 0.030281735584139824, + 0.005105895455926657, + 0.004632466938346624, + 0.004197669681161642, + 0.002676395233720541, + 0.0028827826026827097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.03301645442843437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134400844573975 + ], + "opposite_logprob_tokens": [ + -0.3134400248527527 + ], + "logprob_reward": -1.3134400844573975, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134400844573975, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.023343143984675407, + 0.049664709717035294, + 0.17200765013694763, + 0.35358738899230957, + 0.10529572516679764, + 0.0626259297132492, + 0.07026028633117676, + 0.042080823332071304, + 0.022025667130947113, + 0.09910867363214493 + ], + "credit_map": [ + 0.0035744598135352135, + 0.0025825726334005594, + 0.0024448344483971596, + 0.0030692352447658777, + 0.0023186863400042057, + 0.0037028887309134007, + 0.0024189057294279337, + 0.003231561742722988, + 0.005072473548352718, + 0.004481254145503044, + 0.009229977615177631, + 0.006534372456371784, + 0.0036822138354182243, + 0.007525401655584574, + 0.006357700563967228, + 0.006781314965337515, + 0.01582827791571617, + 0.01618078537285328, + 0.02416793443262577, + 0.02776401676237583, + 0.012556259520351887, + 0.027514969930052757, + 0.021828820928931236, + 0.026166584342718124, + 0.020300395786762238, + 0.09769412875175476, + 0.14118212461471558, + 0.022260740399360657, + 0.014112413860857487, + 0.03388494625687599, + 0.012262580916285515, + 0.01189005933701992, + 0.011982522904872894, + 0.0236466396600008, + 0.014378630556166172, + 0.013740160502493382, + 0.009849824011325836, + 0.011251837015151978, + 0.009409265592694283, + 0.011036847718060017, + 0.004272243473678827, + 0.013334282673895359, + 0.02059607394039631, + 0.006322537548840046, + 0.004564492963254452, + 0.005504576023668051, + 0.0042486595921218395, + 0.003783061634749174, + 0.0069739604368805885, + 0.016799436882138252, + 0.007556335534900427, + 0.009795859456062317, + 0.006491529289633036, + 0.011575080454349518, + 0.0069666458293795586, + 0.004101440776139498, + 0.00443412596359849, + 0.006088662892580032, + 0.00521791772916913, + 0.0038132385816425085, + 0.0046406337060034275, + 0.004493531305342913, + 0.004636446014046669, + 0.008756265975534916, + 0.0015213473234325647, + 0.00507228821516037, + 0.006778974086046219, + 0.0020208843052387238, + 0.0019376745913177729, + 0.001928497338667512, + 0.0015864685410633683, + 0.0011795329628512263, + 0.007063653320074081, + 0.02675558067858219, + 0.04045587405562401, + 0.006614787969738245, + 0.005886733997613192, + 0.005694007035344839, + 0.0028938190080225468, + 0.003744211047887802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -1.3134400844573975 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646786212921143 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008646786212921143, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646786212921143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05804341658949852, + 0.0410078689455986, + 0.04896499216556549, + 0.14496201276779175, + 0.03769529238343239, + 0.2725584805011749, + 0.09750387072563171, + 0.030343035236001015, + 0.042449794709682465, + 0.22647126019001007 + ], + "credit_map": [ + 0.008403433486819267, + 0.004839714150875807, + 0.0074632056057453156, + 0.006285398732870817, + 0.004549517761915922, + 0.009549880400300026, + 0.00711276289075613, + 0.009839503094553947, + 0.0037171763833612204, + 0.004289640113711357, + 0.0059515004977583885, + 0.007899480871856213, + 0.003529266221448779, + 0.006084144581109285, + 0.004594241734594107, + 0.004942420404404402, + 0.005010741762816906, + 0.004641755949705839, + 0.006148409564048052, + 0.007159994915127754, + 0.0036201518960297108, + 0.01030357088893652, + 0.005855098832398653, + 0.00622526602819562, + 0.009025977924466133, + 0.037788450717926025, + 0.04934721440076828, + 0.014356786385178566, + 0.00864502601325512, + 0.010794549249112606, + 0.005848455708473921, + 0.009155545383691788, + 0.004862373229116201, + 0.00596992252394557, + 0.005945349112153053, + 0.0036020150873810053, + 0.0037656831555068493, + 0.0047944700345396996, + 0.004929521586745977, + 0.0038259581197053194, + 0.019537393003702164, + 0.06677858531475067, + 0.10197830945253372, + 0.020392101258039474, + 0.01081576757133007, + 0.01973159797489643, + 0.011048811487853527, + 0.022275887429714203, + 0.0057937586680054665, + 0.017410991713404655, + 0.01669531874358654, + 0.014425489120185375, + 0.01299481000751257, + 0.008015487343072891, + 0.011321371421217918, + 0.01084664836525917, + 0.00285190436989069, + 0.004260040819644928, + 0.004378813318908215, + 0.002924545668065548, + 0.003912791609764099, + 0.004657106939703226, + 0.0028158610220998526, + 0.004541970789432526, + 0.0030176322907209396, + 0.010184390470385551, + 0.013197889551520348, + 0.0036569631192833185, + 0.003163553774356842, + 0.0038503920659422874, + 0.003180048195645213, + 0.002198924543336034, + 0.016273748129606247, + 0.060359321534633636, + 0.09150136262178421, + 0.015451209619641304, + 0.013728704303503036, + 0.013575403019785881, + 0.00719224289059639, + 0.008389277383685112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.008646786212921143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980167791247368 + ], + "opposite_logprob_tokens": [ + -3.529801607131958 + ], + "logprob_reward": -0.02980167791247368, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980167791247368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0427154116332531, + 0.04843023046851158, + 0.019600391387939453, + 0.4664749205112457, + 0.036552198231220245, + 0.05970931798219681, + 0.044180870056152344, + 0.06186067312955856, + 0.025139370933175087, + 0.19533666968345642 + ], + "credit_map": [ + 0.005296520888805389, + 0.003204839304089546, + 0.0071674734354019165, + 0.006109051872044802, + 0.003318917006254196, + 0.005843038205057383, + 0.004236174281686544, + 0.007539394311606884, + 0.004940385930240154, + 0.00294058327563107, + 0.010446594096720219, + 0.010901090689003468, + 0.0020488579757511616, + 0.007105288561433554, + 0.0044643208384513855, + 0.0055831074714660645, + 0.0017240954330191016, + 0.0019639884121716022, + 0.002162342192605138, + 0.0025682360865175724, + 0.00252523529343307, + 0.002560451626777649, + 0.003179772524163127, + 0.002916271099820733, + 0.026118237525224686, + 0.1548464000225067, + 0.17917875945568085, + 0.02954215556383133, + 0.0158955380320549, + 0.03487922623753548, + 0.01211393903940916, + 0.013900671154260635, + 0.0035621076822280884, + 0.005788322538137436, + 0.005535643547773361, + 0.004560970235615969, + 0.004354730248451233, + 0.003648224053904414, + 0.004268703982234001, + 0.004833497107028961, + 0.0038306824862957, + 0.014725456945598125, + 0.021729309111833572, + 0.005152635741978884, + 0.002672443864867091, + 0.00426285108551383, + 0.0031285779550671577, + 0.004207364283502102, + 0.0027665416710078716, + 0.0053808437660336494, + 0.008020750246942043, + 0.005685401149094105, + 0.0067989579401910305, + 0.004877784289419651, + 0.004857323132455349, + 0.005793268792331219, + 0.00466432049870491, + 0.012100040912628174, + 0.021993642672896385, + 0.005010291002690792, + 0.00409028772264719, + 0.005814509931951761, + 0.0039037803653627634, + 0.004283795598894358, + 0.0012266236590221524, + 0.00347731146030128, + 0.003298854222521186, + 0.0033808965235948563, + 0.0033716512843966484, + 0.0034912575501948595, + 0.0039941854774951935, + 0.00289859133772552, + 0.013580619357526302, + 0.05363497510552406, + 0.0781271830201149, + 0.01330606546252966, + 0.011463691480457783, + 0.011378849856555462, + 0.006666958797723055, + 0.007178321480751038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.02980167791247368 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot ultra-wide-angle and taken from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera held still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a blazing meteor visible moving from one corner of the frame to the other?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the meteor's tail transition from brilliant gold near the core to deep blues and purples at the end?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in high-frame-rate slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871236741542816 + ], + "opposite_logprob_tokens": [ + -1.137123703956604 + ], + "logprob_reward": -0.3871236741542816, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871236741542816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04051671177148819, + 0.03013942949473858, + 0.08675491809844971, + 0.15606428682804108, + 0.13565155863761902, + 0.06379453092813492, + 0.0885348692536354, + 0.06172169744968414, + 0.08767925947904587, + 0.24914269149303436 + ], + "credit_map": [ + 0.0055890921503305435, + 0.003424406750127673, + 0.007042362354695797, + 0.005654152482748032, + 0.002818542066961527, + 0.006258219480514526, + 0.003934680484235287, + 0.005795258097350597, + 0.003229630645364523, + 0.002529305638745427, + 0.005167413502931595, + 0.00476850988343358, + 0.0016759909922257066, + 0.0042469697073102, + 0.003953737672418356, + 0.004567871801555157, + 0.009318572469055653, + 0.007741679437458515, + 0.013311098329722881, + 0.011230174452066422, + 0.0065284306183457375, + 0.012916946783661842, + 0.011714274063706398, + 0.01399375032633543, + 0.009503817185759544, + 0.03195016458630562, + 0.03828880563378334, + 0.017375430092215538, + 0.012370430864393711, + 0.020922880619764328, + 0.009356185793876648, + 0.01629658043384552, + 0.009211954660713673, + 0.026748748496174812, + 0.015857337042689323, + 0.019392045214772224, + 0.014738408848643303, + 0.015105178579688072, + 0.01378209050744772, + 0.020815808326005936, + 0.003665103344246745, + 0.014687029644846916, + 0.021692465990781784, + 0.0064447191543877125, + 0.0039726379327476025, + 0.00492436857894063, + 0.00435481546446681, + 0.004053396638482809, + 0.011869416572153568, + 0.01553110871464014, + 0.01664283126592636, + 0.010379120707511902, + 0.009031220339238644, + 0.011172027327120304, + 0.0073454235680401325, + 0.006563724018633366, + 0.0040165879763662815, + 0.012505360879004002, + 0.013432716019451618, + 0.006536444183439016, + 0.0064550163224339485, + 0.0063074128702282906, + 0.006783072371035814, + 0.005685089156031609, + 0.005717411637306213, + 0.0213078111410141, + 0.030382052063941956, + 0.006815499160438776, + 0.006356386002153158, + 0.005832959897816181, + 0.006778726354241371, + 0.004488417878746986, + 0.01786551997065544, + 0.06742900609970093, + 0.10069458186626434, + 0.01704435981810093, + 0.014893255196511745, + 0.014500495977699757, + 0.00784139521420002, + 0.008874109946191311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.3871236741542816 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_6/index_tensor([6955, 6955])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_6/index_tensor([6955, 6955])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..898abc90cf48b4ade530fc4ba7fe89e7cd10f7a4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_6/index_tensor([6955, 6955])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme macro close-up, captured in Annie Leibovitz's signature austere portrait style, the camera remains motionless, darting incredibly close to the face of a figure half-lying in bed. The figure has just cried, and a soft, directed studio light cast from above and to the side precisely captures the traces of tears still lingering at the corners of their eyes. The wet trails create sharp, soft highlights on their skin, causing a few eyelashes to cling together with a slight dampness. The lens is sharply focused on a slender, transparent medical plastic tube, which conforms softly to the contours of the skin. The end of the tube is gently inserted into one of the figure's nostrils, seemingly delivering a barely perceptible stream of air. The tube's walls appear subtly textured in the light. The extreme close-up reveals the figure's skin texture and the finest pores, while the facial muscles appear relaxed and weary after weeping. The background is a completely out-of-focus, soft white bed sheet, forming a pure and blurred block of color, which draws all visual focus to this core detail full of fragility and vitality. The whole picture conveys an atmosphere of profound intimacy, sympathy and human dignity in silence.", + "seed": 19656, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_6/index_6955_seed_19656_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18389823362231256, + "vlm_logprob": -0.18389823362231256, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.18389823362231256, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06273035705089569 + ], + "opposite_logprob_tokens": [ + -3.062730312347412 + ], + "logprob_reward": -0.06273035705089569, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06273035705089569, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5056479573249817, + 0.10008291155099869, + 0.10023325681686401, + 0.04506908729672432, + 0.02819148078560829, + 0.09557133913040161, + 0.040443625301122665, + 0.049007218331098557, + 0.025476474314928055, + 0.010276720859110355 + ], + "credit_map": [ + 0.0307574775069952, + 0.032549917697906494, + 0.050178784877061844, + 0.05693119764328003, + 0.04061434417963028, + 0.03165213018655777, + 0.14635097980499268, + 0.11661306768655777, + 0.006798413582146168, + 0.008395825512707233, + 0.00811090786010027, + 0.009165006689727306, + 0.01690325140953064, + 0.0079896729439497, + 0.02388935536146164, + 0.018830468878149986, + 0.007707439363002777, + 0.005796981044113636, + 0.010231380350887775, + 0.011161652393639088, + 0.01055356115102768, + 0.008096428588032722, + 0.024810124188661575, + 0.02187568135559559, + 0.003256344236433506, + 0.003251332091167569, + 0.004448657855391502, + 0.004166980739682913, + 0.007877719588577747, + 0.0036401741672307253, + 0.009379037655889988, + 0.00904883723706007, + 0.0020452546887099743, + 0.0019461022457107902, + 0.003788679139688611, + 0.004614450037479401, + 0.004568235948681831, + 0.002324118744581938, + 0.004865189548581839, + 0.004039446823298931, + 0.007128831930458546, + 0.008516273461282253, + 0.010207061655819416, + 0.015984496101737022, + 0.009882255457341671, + 0.007361400406807661, + 0.02167346514761448, + 0.014817544259130955, + 0.0032712435349822044, + 0.0035857283510267735, + 0.0052698589861392975, + 0.008882658556103706, + 0.003088300349190831, + 0.002784541342407465, + 0.008402533829212189, + 0.005158756859600544, + 0.0036462554708123207, + 0.004830243997275829, + 0.004294620361179113, + 0.006719944998621941, + 0.00951700285077095, + 0.004426965024322271, + 0.008951768279075623, + 0.006620411295443773, + 0.0016498927725479007, + 0.002268986077979207, + 0.0020267243962734938, + 0.002691758330911398, + 0.003899000119417906, + 0.002099031815305352, + 0.006210808642208576, + 0.004630270414054394, + 0.0006872476660646498, + 0.0006373150390572846, + 0.0007556953933089972, + 0.001038425718434155, + 0.0010230124462395906, + 0.0009249777067452669, + 0.003095319727435708, + 0.002114726696163416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.06273035705089569 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934124827384949 + ], + "opposite_logprob_tokens": [ + -0.6934124827384949 + ], + "logprob_reward": -0.6934124827384949, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934124827384949, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.42505645751953125, + 0.08475527167320251, + 0.10377121716737747, + 0.05726601555943489, + 0.071353018283844, + 0.07564299553632736, + 0.04151325300335884, + 0.056998442858457565, + 0.040407050400972366, + 0.04323623329401016 + ], + "credit_map": [ + 0.03626776114106178, + 0.027404360473155975, + 0.059290915727615356, + 0.06796462088823318, + 0.029976755380630493, + 0.02552369423210621, + 0.10440309345722198, + 0.07422523945569992, + 0.0051053306087851524, + 0.005532271694391966, + 0.0066202604211866856, + 0.010279507376253605, + 0.012091001495718956, + 0.007514096796512604, + 0.024137325584888458, + 0.013475474901497364, + 0.006514746230095625, + 0.00589537201449275, + 0.008808541111648083, + 0.011881466023623943, + 0.012633276171982288, + 0.00661902641877532, + 0.029752561822533607, + 0.021666226908564568, + 0.005671393126249313, + 0.004219936206936836, + 0.009097469039261341, + 0.008516578003764153, + 0.007814722135663033, + 0.004720160271972418, + 0.008393784984946251, + 0.008831971324980259, + 0.005521421320736408, + 0.006782147567719221, + 0.008226189762353897, + 0.012033762410283089, + 0.008332542143762112, + 0.005831466056406498, + 0.013538156636059284, + 0.0110873281955719, + 0.006309165619313717, + 0.007790705189108849, + 0.006946277339011431, + 0.007995735853910446, + 0.018748030066490173, + 0.005235269200056791, + 0.013720426708459854, + 0.008897383697330952, + 0.0039548990316689014, + 0.004162677563726902, + 0.0041891406290233135, + 0.004388295114040375, + 0.008288798853754997, + 0.002865080488845706, + 0.005899430718272924, + 0.007764935027807951, + 0.005827085580676794, + 0.005644309800118208, + 0.007844987325370312, + 0.012386824935674667, + 0.008539075963199139, + 0.0037517023738473654, + 0.00662032188847661, + 0.006384130567312241, + 0.0038647593464702368, + 0.004511629696935415, + 0.003739073406904936, + 0.0044579156674444675, + 0.007570877205580473, + 0.0035238945856690407, + 0.007140521425753832, + 0.005598379299044609, + 0.0024994048289954662, + 0.003684448776766658, + 0.00318762450478971, + 0.004623296670615673, + 0.006908361800014973, + 0.003088186262175441, + 0.01069049071520567, + 0.008554416708648205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.6934124827384949 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588496699929237 + ], + "opposite_logprob_tokens": [ + -3.8955883979797363 + ], + "logprob_reward": -0.020588496699929237, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588496699929237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36500757932662964, + 0.1362391710281372, + 0.1789180487394333, + 0.047375891357660294, + 0.07326508313417435, + 0.060337625443935394, + 0.03180006891489029, + 0.05041038617491722, + 0.023886103183031082, + 0.03276000916957855 + ], + "credit_map": [ + 0.023103831335902214, + 0.022236445918679237, + 0.04095175862312317, + 0.042113784700632095, + 0.03983018547296524, + 0.02591032348573208, + 0.09669455140829086, + 0.07416673004627228, + 0.009999232366681099, + 0.012923652306199074, + 0.011963403783738613, + 0.014664649032056332, + 0.028209138661623, + 0.012864457443356514, + 0.026317767798900604, + 0.019296864047646523, + 0.01013974566012621, + 0.01244604866951704, + 0.01596122793853283, + 0.017853861674666405, + 0.026263311505317688, + 0.011578286066651344, + 0.04827352240681648, + 0.036402057856321335, + 0.0043286788277328014, + 0.004101887810975313, + 0.006894109304994345, + 0.006437886040657759, + 0.007487942930310965, + 0.00486746896058321, + 0.006228276994079351, + 0.007029641419649124, + 0.005585141014307737, + 0.00827875081449747, + 0.009473135694861412, + 0.011258944869041443, + 0.013631868176162243, + 0.005558617878705263, + 0.009670570492744446, + 0.009808055125176907, + 0.005379826761782169, + 0.007692118175327778, + 0.005417390260845423, + 0.0073060765862464905, + 0.01593044586479664, + 0.003991090226918459, + 0.007910226471722126, + 0.006710449233651161, + 0.002667010296136141, + 0.0032793511636555195, + 0.0037842928431928158, + 0.004274277947843075, + 0.006206400692462921, + 0.0028040530160069466, + 0.003506589215248823, + 0.005278091412037611, + 0.004711995366960764, + 0.005218483507633209, + 0.006704260129481554, + 0.009967811405658722, + 0.007013730704784393, + 0.0038994289934635162, + 0.0064729684963822365, + 0.0064217038452625275, + 0.001624662079848349, + 0.002193290973082185, + 0.0023898088838905096, + 0.0025240795221179724, + 0.0027884014416486025, + 0.0022979644127190113, + 0.004675691481679678, + 0.005392203573137522, + 0.002734706038609147, + 0.0033966735936701298, + 0.0025555193424224854, + 0.0023413116578012705, + 0.00864492729306221, + 0.004028052557259798, + 0.0047825356014072895, + 0.004276284482330084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.020588496699929237 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296301335096359 + ], + "opposite_logprob_tokens": [ + -2.2379629611968994 + ], + "logprob_reward": -0.11296301335096359, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296301335096359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47330209612846375, + 0.17995040118694305, + 0.049637869000434875, + 0.07632317394018173, + 0.08013845980167389, + 0.04760708659887314, + 0.018840448930859566, + 0.027898460626602173, + 0.028181282803416252, + 0.01812072843313217 + ], + "credit_map": [ + 0.02964743971824646, + 0.0316624753177166, + 0.04395986720919609, + 0.05018257722258568, + 0.06002921238541603, + 0.03142671659588814, + 0.12474250048398972, + 0.10165129601955414, + 0.013095226138830185, + 0.014124272391200066, + 0.015421966090798378, + 0.015878604725003242, + 0.03262779116630554, + 0.013694565743207932, + 0.040284041315317154, + 0.034823939204216, + 0.00438185827806592, + 0.0043548294343054295, + 0.006261512637138367, + 0.008097662590444088, + 0.008655350655317307, + 0.004721776582300663, + 0.007043118122965097, + 0.006121760234236717, + 0.007081545423716307, + 0.00594515074044466, + 0.010211220942437649, + 0.009373048320412636, + 0.016566291451454163, + 0.005931437481194735, + 0.010090097784996033, + 0.011124386452138424, + 0.0072924490086734295, + 0.006690671667456627, + 0.008745528757572174, + 0.012379988096654415, + 0.008143291808664799, + 0.006705031730234623, + 0.018493006005883217, + 0.011688491329550743, + 0.0042803636752069, + 0.004387739580124617, + 0.006179959513247013, + 0.008669697679579258, + 0.006816365756094456, + 0.003689769422635436, + 0.00824073888361454, + 0.005342455115169287, + 0.0016646081348881125, + 0.001143507775850594, + 0.0020784635562449694, + 0.0024551423266530037, + 0.002682626945897937, + 0.0016352348029613495, + 0.0034542453940957785, + 0.0037266206927597523, + 0.0024430251214653254, + 0.002973172813653946, + 0.002472996711730957, + 0.0029360312037169933, + 0.008624419569969177, + 0.002549047814682126, + 0.0028333610389381647, + 0.003066406585276127, + 0.002217016415670514, + 0.0025801670271903276, + 0.0026770150288939476, + 0.0023814321029931307, + 0.005177662707865238, + 0.0027280228678137064, + 0.004344808403402567, + 0.006075158715248108, + 0.0012345756404101849, + 0.0014461760874837637, + 0.0017668266082182527, + 0.0019543154630810022, + 0.0014737684978172183, + 0.001458129147067666, + 0.004793549422174692, + 0.003993387334048748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.11296301335096359 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796818271279335 + ], + "opposite_logprob_tokens": [ + -3.529796838760376 + ], + "logprob_reward": -0.029796818271279335, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796818271279335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37913089990615845, + 0.1719481199979782, + 0.11605353653430939, + 0.026077480986714363, + 0.09137800335884094, + 0.04829142615199089, + 0.050695594400167465, + 0.05482853949069977, + 0.029904203489422798, + 0.031692154705524445 + ], + "credit_map": [ + 0.03213740512728691, + 0.041566651314496994, + 0.043683238327503204, + 0.049610745161771774, + 0.07296574860811234, + 0.030402138829231262, + 0.054846882820129395, + 0.05391809344291687, + 0.010269354097545147, + 0.013600672595202923, + 0.014453780837357044, + 0.02459966391324997, + 0.02117290161550045, + 0.014488287270069122, + 0.04420599341392517, + 0.029157456010580063, + 0.007294945418834686, + 0.008387577719986439, + 0.012800627388060093, + 0.016858326271176338, + 0.008253115229308605, + 0.008023654110729694, + 0.028801394626498222, + 0.02563389018177986, + 0.0024079792201519012, + 0.0026448690332472324, + 0.0036825230345129967, + 0.0036017312668263912, + 0.0034210237208753824, + 0.0029301936738193035, + 0.003088707569986582, + 0.0043004523031413555, + 0.008510379120707512, + 0.01044519804418087, + 0.01491846889257431, + 0.0192559901624918, + 0.013612700626254082, + 0.006570529192686081, + 0.008520080707967281, + 0.009544648230075836, + 0.005319401156157255, + 0.004415852017700672, + 0.007491393946111202, + 0.010408707894384861, + 0.006152155809104443, + 0.004238951485604048, + 0.004686335567384958, + 0.005578627344220877, + 0.004560139495879412, + 0.006465002428740263, + 0.00436886353418231, + 0.004838923458009958, + 0.01489891204982996, + 0.004295348655432463, + 0.005381109192967415, + 0.005887292791157961, + 0.004089527763426304, + 0.004941615741699934, + 0.005105702672153711, + 0.007471901830285788, + 0.008651886135339737, + 0.005086924880743027, + 0.010343200527131557, + 0.009137777611613274, + 0.0031529210973531008, + 0.0035068506840616465, + 0.0029515735805034637, + 0.004546103533357382, + 0.005698959343135357, + 0.0021951848175376654, + 0.0044490075670182705, + 0.0034036021679639816, + 0.0031518384348601103, + 0.0036809586454182863, + 0.00437250966206193, + 0.004770671483129263, + 0.005250471644103527, + 0.002420412376523018, + 0.004398034419864416, + 0.0036472587380558252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.029796818271279335 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme macro close-up, captured in Annie Leibovitz's signature austere portrait style, the camera remains motionless, darting incredibly close to the face of a figure half-lying in bed. The figure has just cried, and a soft, directed studio light cast from above and to the side precisely captures the traces of tears still lingering at the corners of their eyes. The wet trails create sharp, soft highlights on their skin, causing a few eyelashes to cling together with a slight dampness. The lens is sharply focused on a slender, transparent medical plastic tube, which conforms softly to the contours of the skin. The end of the tube is gently inserted into one of the figure's nostrils, seemingly delivering a barely perceptible stream of air. The tube's walls appear subtly textured in the light. The extreme close-up reveals the figure's skin texture and the finest pores, while the facial muscles appear relaxed and weary after weeping. The background is a completely out-of-focus, soft white bed sheet, forming a pure and blurred block of color, which draws all visual focus to this core detail full of fragility and vitality. The whole picture conveys an atmosphere of profound intimacy, sympathy and human dignity in silence.", + "seed": 19657, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_6/index_6955_seed_19657_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3018162716180086, + "vlm_logprob": -0.3018162716180086, + "vlm_yes_no_margin": 1.75, + "vlm_token_credit": -0.3018162716180086, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11240417510271072 + ], + "opposite_logprob_tokens": [ + -2.3624041080474854 + ], + "logprob_reward": -0.11240417510271072, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11240417510271072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3878462314605713, + 0.0966629758477211, + 0.05842514708638191, + 0.04063110426068306, + 0.014309830032289028, + 0.08823443204164505, + 0.11807549744844437, + 0.05168629437685013, + 0.03689790517091751, + 0.1072305291891098 + ], + "credit_map": [ + 0.04331198334693909, + 0.034015607088804245, + 0.04614170640707016, + 0.06464140862226486, + 0.04124869406223297, + 0.035460133105516434, + 0.08210323750972748, + 0.040923476219177246, + 0.011805669404566288, + 0.007896114140748978, + 0.01128872949630022, + 0.015245248563587666, + 0.013415729627013206, + 0.008528423495590687, + 0.010734044946730137, + 0.01774902641773224, + 0.0059624481946229935, + 0.004788929596543312, + 0.004832049831748009, + 0.009562728926539421, + 0.008054498583078384, + 0.005920166149735451, + 0.008682801388204098, + 0.010621530935168266, + 0.0046517131850123405, + 0.0026693581603467464, + 0.003798993770033121, + 0.005417404696345329, + 0.007847288623452187, + 0.005153912119567394, + 0.005047012586146593, + 0.0060454243794083595, + 0.0013140238588675857, + 0.0011709260288625956, + 0.0012719547376036644, + 0.00218328763730824, + 0.00238111917860806, + 0.0013379270676523447, + 0.002673697890713811, + 0.0019768946804106236, + 0.01132231019437313, + 0.006075213197618723, + 0.010122906416654587, + 0.013375477865338326, + 0.015112032182514668, + 0.008167996071279049, + 0.012356679886579514, + 0.011701817624270916, + 0.01376641821116209, + 0.006493902299553156, + 0.013913562521338463, + 0.02038286067545414, + 0.02066974900662899, + 0.009597284719347954, + 0.019645001739263535, + 0.013606729917228222, + 0.00566976610571146, + 0.005681511014699936, + 0.007152094505727291, + 0.008736561052501202, + 0.004575670231133699, + 0.003760727820917964, + 0.008248561061918736, + 0.00786140188574791, + 0.004137068521231413, + 0.003549330634996295, + 0.0035504086408764124, + 0.0054267155937850475, + 0.0035686257760971785, + 0.003803589614108205, + 0.006024129223078489, + 0.006838039495050907, + 0.021962840110063553, + 0.005088447127491236, + 0.009672597981989384, + 0.009040024131536484, + 0.0066841342486441135, + 0.034062210470438004, + 0.011697324924170971, + 0.00902295671403408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.11240417510271072 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743409752845764 + ], + "opposite_logprob_tokens": [ + -0.4743409752845764 + ], + "logprob_reward": -0.9743409752845764, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743409752845764, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14547553658485413, + 0.06915205717086792, + 0.049889516085386276, + 0.08248649537563324, + 0.09796932339668274, + 0.024648364633321762, + 0.07593973726034164, + 0.13713732361793518, + 0.23154261708259583, + 0.08575901389122009 + ], + "credit_map": [ + 0.016495009884238243, + 0.013555178418755531, + 0.013707603327929974, + 0.024026265367865562, + 0.02274898625910282, + 0.0112399160861969, + 0.02574276551604271, + 0.017959807068109512, + 0.009368584491312504, + 0.0038388874381780624, + 0.011038431897759438, + 0.009827418252825737, + 0.008523778058588505, + 0.006154832895845175, + 0.009155449457466602, + 0.011244666762650013, + 0.0049856286495924, + 0.003517299657687545, + 0.006561522837728262, + 0.008353556506335735, + 0.006024844944477081, + 0.00696717482060194, + 0.0066246166825294495, + 0.0068548633717000484, + 0.009476345032453537, + 0.006803716998547316, + 0.007872753776609898, + 0.012497867457568645, + 0.010548166930675507, + 0.008800022304058075, + 0.010927392169833183, + 0.015560229308903217, + 0.011874249204993248, + 0.007874147966504097, + 0.006884666159749031, + 0.01046742033213377, + 0.02905493974685669, + 0.0089540621265769, + 0.015237518586218357, + 0.0076223150826990604, + 0.0028583407402038574, + 0.0024670567363500595, + 0.0027527266647666693, + 0.003863763064146042, + 0.0036993820685893297, + 0.003400118788704276, + 0.003588547930121422, + 0.002018427476286888, + 0.009018498472869396, + 0.005153616424649954, + 0.006925743538886309, + 0.0112612284719944, + 0.011118386872112751, + 0.007539587095379829, + 0.01232604868710041, + 0.01259662676602602, + 0.015499516390264034, + 0.009570199064910412, + 0.016022400930523872, + 0.020189063623547554, + 0.016609076410531998, + 0.011345197446644306, + 0.023576440289616585, + 0.024325421079993248, + 0.03082606941461563, + 0.031431518495082855, + 0.025467513129115105, + 0.040834661573171616, + 0.02828139252960682, + 0.014969072304666042, + 0.032527267932891846, + 0.027205120772123337, + 0.016631657257676125, + 0.0044113039039075375, + 0.006916207727044821, + 0.008692270144820213, + 0.0059546115808188915, + 0.026658423244953156, + 0.009324179962277412, + 0.00717034749686718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.9743409752845764 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982609160244465 + ], + "opposite_logprob_tokens": [ + -3.5298261642456055 + ], + "logprob_reward": -0.02982609160244465, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982609160244465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19707952439785004, + 0.15570016205310822, + 0.1394219845533371, + 0.17587587237358093, + 0.09218236804008484, + 0.05238357186317444, + 0.07794097065925598, + 0.043056562542915344, + 0.02256813645362854, + 0.04379086196422577 + ], + "credit_map": [ + 0.018673822283744812, + 0.015149528160691261, + 0.018656985834240913, + 0.03497787564992905, + 0.02092750556766987, + 0.02586222067475319, + 0.036514125764369965, + 0.026317443698644638, + 0.013767629861831665, + 0.013891528360545635, + 0.018287090584635735, + 0.021480416879057884, + 0.025043027475476265, + 0.016008632257580757, + 0.021342799067497253, + 0.02587902918457985, + 0.012826773338019848, + 0.01129711139947176, + 0.010861756280064583, + 0.023148903623223305, + 0.018510302528738976, + 0.014344391413033009, + 0.024064742028713226, + 0.024367988109588623, + 0.019468465819954872, + 0.010588310658931732, + 0.013594835996627808, + 0.02222970314323902, + 0.0459575392305851, + 0.02137054316699505, + 0.02244417369365692, + 0.020222293213009834, + 0.01121530681848526, + 0.011778965592384338, + 0.007626333739608526, + 0.010166036896407604, + 0.021904367953538895, + 0.010171725414693356, + 0.0118073346093297, + 0.00751228304579854, + 0.005888352170586586, + 0.003634006716310978, + 0.0049758534878492355, + 0.00847321841865778, + 0.010387702845036983, + 0.005108006298542023, + 0.00743783637881279, + 0.006478589493781328, + 0.007676302455365658, + 0.004747871775180101, + 0.008081424050033092, + 0.012492445297539234, + 0.016658317297697067, + 0.005990692414343357, + 0.011577489785850048, + 0.010716420598328114, + 0.004185310564935207, + 0.003929013852030039, + 0.00538014667108655, + 0.0071544344536960125, + 0.006623571738600731, + 0.0048172553069889545, + 0.0064247166737914085, + 0.004542107693850994, + 0.0031040599569678307, + 0.0018332658801227808, + 0.003570500062778592, + 0.003626161487773061, + 0.0017326846718788147, + 0.0020850806031376123, + 0.0035037442576140165, + 0.0031126353424042463, + 0.007524480111896992, + 0.002417521085590124, + 0.003856007009744644, + 0.005770975723862648, + 0.004099126905202866, + 0.008520863018929958, + 0.0057388548739254475, + 0.005863027647137642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.02982609160244465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135174810886383 + ], + "opposite_logprob_tokens": [ + -1.313517451286316 + ], + "logprob_reward": -0.3135174810886383, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135174810886383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35233479738235474, + 0.14694330096244812, + 0.10217954218387604, + 0.042053718119859695, + 0.01725563034415245, + 0.031593553721904755, + 0.14332133531570435, + 0.04164521023631096, + 0.05517458915710449, + 0.06749832630157471 + ], + "credit_map": [ + 0.03457839787006378, + 0.029984496533870697, + 0.03655487671494484, + 0.06350374221801758, + 0.03945958614349365, + 0.03876862674951553, + 0.06579042971134186, + 0.04369460791349411, + 0.014029473066329956, + 0.011395812034606934, + 0.013462839648127556, + 0.023707345128059387, + 0.02514627017080784, + 0.011527353897690773, + 0.016390299424529076, + 0.0312839038670063, + 0.008882659487426281, + 0.008571826852858067, + 0.007085006218403578, + 0.016022225841879845, + 0.015363475307822227, + 0.009375025518238544, + 0.015727518126368523, + 0.02115180343389511, + 0.004100233316421509, + 0.0031538407783955336, + 0.004307347349822521, + 0.004923682659864426, + 0.006596988067030907, + 0.006819373928010464, + 0.0058599915355443954, + 0.006292259786278009, + 0.002492106519639492, + 0.0016966407420113683, + 0.00217705057002604, + 0.002520293463021517, + 0.0031668951269239187, + 0.0011679811868816614, + 0.002674864372238517, + 0.0013597963843494654, + 0.003937243018299341, + 0.002175865927711129, + 0.0028131187427788973, + 0.003507835790514946, + 0.006831686943769455, + 0.00290428358130157, + 0.005333247594535351, + 0.004090269561856985, + 0.0142985163256526, + 0.00857099611312151, + 0.013108531013131142, + 0.02123541571199894, + 0.02994033321738243, + 0.013891969807446003, + 0.02191345952451229, + 0.02036210708320141, + 0.004406007472425699, + 0.003318128176033497, + 0.005062649492174387, + 0.007522508967667818, + 0.006864612922072411, + 0.0032315354328602552, + 0.0072226813063025475, + 0.004017082042992115, + 0.006454729475080967, + 0.006946056615561247, + 0.004757273476570845, + 0.010472971014678478, + 0.006251529324799776, + 0.004086649511009455, + 0.008169014006853104, + 0.008036359213292599, + 0.01420747209340334, + 0.003291693050414324, + 0.00580605398863554, + 0.005910805426537991, + 0.005106976721435785, + 0.01959381438791752, + 0.00769491633400321, + 0.005886591039597988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.3135174810886383 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899263501167297 + ], + "opposite_logprob_tokens": [ + -2.5789926052093506 + ], + "logprob_reward": -0.07899263501167297, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899263501167297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3682825267314911, + 0.1944650113582611, + 0.11186947673559189, + 0.033640023320913315, + 0.03478648513555527, + 0.05762970447540283, + 0.07493875175714493, + 0.044145870953798294, + 0.043167099356651306, + 0.037075091153383255 + ], + "credit_map": [ + 0.03571320325136185, + 0.033175431191921234, + 0.03753311559557915, + 0.06538063287734985, + 0.036121997982263565, + 0.046247199177742004, + 0.0653235912322998, + 0.04878734052181244, + 0.020675774663686752, + 0.013741008006036282, + 0.03341316804289818, + 0.025174876675009727, + 0.02563122659921646, + 0.021411074325442314, + 0.022621842101216316, + 0.031796034425497055, + 0.014035302214324474, + 0.009789057075977325, + 0.011819588951766491, + 0.0157579705119133, + 0.015237895771861076, + 0.008947499096393585, + 0.021664712578058243, + 0.014617439359426498, + 0.002930111950263381, + 0.0021753746550530195, + 0.002746693091467023, + 0.0049310713075101376, + 0.007667927537113428, + 0.003274638671427965, + 0.004783460404723883, + 0.005130743142217398, + 0.005927226971834898, + 0.00275902496650815, + 0.0041322228498756886, + 0.004513890948146582, + 0.004146633204072714, + 0.002885442692786455, + 0.005586587358266115, + 0.004835451487451792, + 0.005336642265319824, + 0.004671749658882618, + 0.006170371081680059, + 0.01043383777141571, + 0.008503172546625137, + 0.005357412155717611, + 0.008477691560983658, + 0.008678821846842766, + 0.0084840077906847, + 0.0054917968809604645, + 0.010308885015547276, + 0.010395591147243977, + 0.014554280787706375, + 0.006226480007171631, + 0.012150506488978863, + 0.007327195256948471, + 0.005253598093986511, + 0.005903319921344519, + 0.003986836411058903, + 0.007126538082957268, + 0.005954295862466097, + 0.0035726719070225954, + 0.00577537203207612, + 0.006573233753442764, + 0.005097819026559591, + 0.0059525431133806705, + 0.005287473089993, + 0.007400777190923691, + 0.005206695757806301, + 0.002740431809797883, + 0.0059341504238545895, + 0.005547208245843649, + 0.005627933423966169, + 0.0027141927275806665, + 0.0034721032716333866, + 0.004852228332310915, + 0.00763175543397665, + 0.0032020441722124815, + 0.005108773708343506, + 0.004466052632778883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07899263501167297 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme macro close-up, captured in Annie Leibovitz's signature austere portrait style, the camera remains motionless, darting incredibly close to the face of a figure half-lying in bed. The figure has just cried, and a soft, directed studio light cast from above and to the side precisely captures the traces of tears still lingering at the corners of their eyes. The wet trails create sharp, soft highlights on their skin, causing a few eyelashes to cling together with a slight dampness. The lens is sharply focused on a slender, transparent medical plastic tube, which conforms softly to the contours of the skin. The end of the tube is gently inserted into one of the figure's nostrils, seemingly delivering a barely perceptible stream of air. The tube's walls appear subtly textured in the light. The extreme close-up reveals the figure's skin texture and the finest pores, while the facial muscles appear relaxed and weary after weeping. The background is a completely out-of-focus, soft white bed sheet, forming a pure and blurred block of color, which draws all visual focus to this core detail full of fragility and vitality. The whole picture conveys an atmosphere of profound intimacy, sympathy and human dignity in silence.", + "seed": 19658, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_6/index_6955_seed_19658_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.31564853712916374, + "vlm_logprob": -0.31564853712916374, + "vlm_yes_no_margin": 1.575, + "vlm_token_credit": -0.31564853712916374, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11116792261600494 + ], + "opposite_logprob_tokens": [ + -2.3611679077148438 + ], + "logprob_reward": -0.11116792261600494, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11116792261600494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09588994830846786, + 0.2042151838541031, + 0.02089676633477211, + 0.019864728674292564, + 0.5004719495773315, + 0.06883317977190018, + 0.0182966236025095, + 0.013335119001567364, + 0.032839078456163406, + 0.025357414036989212 + ], + "credit_map": [ + 0.004508723970502615, + 0.004572154022753239, + 0.005424644332379103, + 0.006139615550637245, + 0.0069426060654222965, + 0.03110136277973652, + 0.007258187979459763, + 0.02994265966117382, + 0.007193943951278925, + 0.00358198513276875, + 0.009105850011110306, + 0.012407341040670872, + 0.008240344934165478, + 0.07371741533279419, + 0.009177002124488354, + 0.08079131692647934, + 0.001771357492543757, + 0.0015473481034860015, + 0.002499973401427269, + 0.0021119641605764627, + 0.002007140079513192, + 0.006043056957423687, + 0.002021387917920947, + 0.002894539386034012, + 0.001393684884533286, + 0.00145884754601866, + 0.004757678136229515, + 0.0023919488303363323, + 0.001248618122190237, + 0.0038512363098561764, + 0.002434986177831888, + 0.002327726921066642, + 0.02736085280776024, + 0.014179966412484646, + 0.16482563316822052, + 0.048360031098127365, + 0.03148223087191582, + 0.16774025559425354, + 0.029040910303592682, + 0.017482034862041473, + 0.004127010703086853, + 0.003671613521873951, + 0.019686240702867508, + 0.007264524698257446, + 0.004249830264598131, + 0.0177390668541193, + 0.006611790042370558, + 0.00548310624435544, + 0.0014808685518801212, + 0.0012169857509434223, + 0.0035241623409092426, + 0.002209843136370182, + 0.0015849883202463388, + 0.0026886281557381153, + 0.0030775018967688084, + 0.0025136445183306932, + 0.0014966384042054415, + 0.0010767672210931778, + 0.002939992118626833, + 0.0010796780698001385, + 0.0006610443233512342, + 0.0018659873167052865, + 0.0010053926380351186, + 0.0032096179202198982, + 0.002377294236794114, + 0.0032792144920676947, + 0.002338733524084091, + 0.0028052516281604767, + 0.0025997364427894354, + 0.007602940779179335, + 0.0030517575796693563, + 0.008784148842096329, + 0.0034001809544861317, + 0.002640503691509366, + 0.0036250220146030188, + 0.0018978920998051763, + 0.0012861285358667374, + 0.0025750065688043833, + 0.002195797860622406, + 0.00773688405752182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.11116792261600494 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743366837501526 + ], + "opposite_logprob_tokens": [ + -0.4743366837501526 + ], + "logprob_reward": -0.9743366837501526, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743366837501526, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04593641683459282, + 0.07558662444353104, + 0.027941742911934853, + 0.11070975661277771, + 0.5217472314834595, + 0.048434823751449585, + 0.03496769443154335, + 0.045443400740623474, + 0.0663098692893982, + 0.022922461852431297 + ], + "credit_map": [ + 0.004275819286704063, + 0.004314382094889879, + 0.005937568843364716, + 0.005602353718131781, + 0.004116213880479336, + 0.006380562670528889, + 0.006689496338367462, + 0.008620020933449268, + 0.0045476569794118404, + 0.005246151704341173, + 0.009087955579161644, + 0.009045039303600788, + 0.004782423377037048, + 0.01450376957654953, + 0.01049231830984354, + 0.01788131147623062, + 0.002713780850172043, + 0.0018594349967315793, + 0.0035817136522382498, + 0.0033785505220294, + 0.0028854189440608025, + 0.003261965699493885, + 0.005056446883827448, + 0.005204429384320974, + 0.00961981900036335, + 0.008070278912782669, + 0.02479604445397854, + 0.01302334014326334, + 0.006787386257201433, + 0.014032297767698765, + 0.013871480710804462, + 0.020509107038378716, + 0.04095286503434181, + 0.030594831332564354, + 0.15462978184223175, + 0.04597342386841774, + 0.03188151493668556, + 0.16196873784065247, + 0.028790730983018875, + 0.02695532888174057, + 0.0031233918853104115, + 0.0027970559895038605, + 0.014064675197005272, + 0.005048365797847509, + 0.0028203376568853855, + 0.009993143379688263, + 0.0046999393962323666, + 0.005887910723686218, + 0.003921261988580227, + 0.0025349443312734365, + 0.007417921908199787, + 0.004003091715276241, + 0.003005868522450328, + 0.003980310168117285, + 0.00526069151237607, + 0.004843604750931263, + 0.002719347830861807, + 0.0023874000180512667, + 0.008102054707705975, + 0.003406747244298458, + 0.002594170393422246, + 0.008869431912899017, + 0.003755483776330948, + 0.013608764857053757, + 0.005123654380440712, + 0.003685510251671076, + 0.009947053156793118, + 0.005016035865992308, + 0.003840600373223424, + 0.012756879441440105, + 0.005504992790520191, + 0.020435141399502754, + 0.0021934444084763527, + 0.0014096855884417892, + 0.0023375290911644697, + 0.0015654648886993527, + 0.0016992660239338875, + 0.0046529644168913364, + 0.002136759925633669, + 0.006927348207682371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.9743366837501526 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054989952594041824 + ], + "opposite_logprob_tokens": [ + -2.92999005317688 + ], + "logprob_reward": -0.054989952594041824, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054989952594041824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19464661180973053, + 0.33436301350593567, + 0.03720240667462349, + 0.02189839445054531, + 0.25179705023765564, + 0.03528522700071335, + 0.010843390598893166, + 0.03580500930547714, + 0.04320027306675911, + 0.03495858609676361 + ], + "credit_map": [ + 0.00911642424762249, + 0.007627180311828852, + 0.010745364241302013, + 0.014027084223926067, + 0.012824593111872673, + 0.062408026307821274, + 0.01660127379000187, + 0.061296671628952026, + 0.011896046809852123, + 0.007739691529422998, + 0.01524329837411642, + 0.020834898576140404, + 0.014188653789460659, + 0.11914006620645523, + 0.016903454437851906, + 0.12841691076755524, + 0.0042423708364367485, + 0.003363144351169467, + 0.004232791718095541, + 0.0050958674401044846, + 0.003479328704997897, + 0.004650781396776438, + 0.006144727114588022, + 0.005993396043777466, + 0.0020667738281190395, + 0.002448358340188861, + 0.003307194449007511, + 0.0025714077055454254, + 0.0026710201054811478, + 0.0025825267657637596, + 0.0035544505808502436, + 0.002696661977097392, + 0.014596044085919857, + 0.009111094288527966, + 0.08072174340486526, + 0.02379429154098034, + 0.015853900462388992, + 0.08464086800813675, + 0.014218599535524845, + 0.008860522881150246, + 0.0020430744625627995, + 0.0018716634949669242, + 0.009588046930730343, + 0.0033288784325122833, + 0.0024998807348310947, + 0.009126953780651093, + 0.0038707612548023462, + 0.0029559703543782234, + 0.001680832589045167, + 0.0011583180166780949, + 0.000988660380244255, + 0.0010532874148339033, + 0.0014453165931627154, + 0.0017059085657820106, + 0.0013872248819097877, + 0.0014238432049751282, + 0.0017451124731451273, + 0.0014459611847996712, + 0.004146479070186615, + 0.002019366016611457, + 0.0018104983028024435, + 0.010800384916365147, + 0.002271382138133049, + 0.01156582497060299, + 0.002788752783089876, + 0.003922303672879934, + 0.0027874093502759933, + 0.0032319563906639814, + 0.004046908114105463, + 0.010226000100374222, + 0.005195159465074539, + 0.011001783423125744, + 0.002042690757662058, + 0.002765315119177103, + 0.0025289778131991625, + 0.003739019390195608, + 0.0034675297793000937, + 0.006746769417077303, + 0.006987939588725567, + 0.006680341903120279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.054989952594041824 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34870266914367676 + ], + "opposite_logprob_tokens": [ + -1.2237026691436768 + ], + "logprob_reward": -0.34870266914367676, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34870266914367676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06403200328350067, + 0.24265989661216736, + 0.022639794275164604, + 0.05621920898556709, + 0.42626330256462097, + 0.061695560812950134, + 0.03532000631093979, + 0.014211884699761868, + 0.03090314194560051, + 0.04605521261692047 + ], + "credit_map": [ + 0.0065607656724750996, + 0.009267211891710758, + 0.007563238963484764, + 0.006454394198954105, + 0.007824871689081192, + 0.006623527035117149, + 0.011246241629123688, + 0.00849175825715065, + 0.00926730502396822, + 0.005933987908065319, + 0.013632120564579964, + 0.015896618366241455, + 0.010128635913133621, + 0.08139392733573914, + 0.015319359488785267, + 0.0910879373550415, + 0.0032777306623756886, + 0.0024334415793418884, + 0.0027821436524391174, + 0.0026075912173837423, + 0.0022010700777173042, + 0.0025635522324591875, + 0.0026437260676175356, + 0.004130537621676922, + 0.007603941019624472, + 0.004470855463296175, + 0.012279829941689968, + 0.005439899396151304, + 0.0033624584320932627, + 0.008574520237743855, + 0.006086284294724464, + 0.008401423692703247, + 0.023487424477934837, + 0.011660286225378513, + 0.1405389904975891, + 0.04118999466300011, + 0.026712963357567787, + 0.14174926280975342, + 0.02516731433570385, + 0.015757082030177116, + 0.004274084698408842, + 0.0045592887327075005, + 0.013769765384495258, + 0.006863194052129984, + 0.0038498181384056807, + 0.014808690175414085, + 0.007162659894675016, + 0.00640806183218956, + 0.002458961447700858, + 0.002405203180387616, + 0.006576712243258953, + 0.0038974343333393335, + 0.0029868916608393192, + 0.005728446412831545, + 0.005935286171734333, + 0.0053310696966946125, + 0.0015741820679977536, + 0.001450825366191566, + 0.0029006311669945717, + 0.0010886875679716468, + 0.0010505507234483957, + 0.0011927797459065914, + 0.00170931126922369, + 0.003244916209951043, + 0.003568848827853799, + 0.0033735064789652824, + 0.002445767167955637, + 0.0032223218586295843, + 0.0019001258770003915, + 0.006546385120600462, + 0.0027994492556899786, + 0.0070467377081513405, + 0.003411472775042057, + 0.0030287632253021, + 0.005354901775717735, + 0.003351801773533225, + 0.0037707178853452206, + 0.007792122662067413, + 0.004941998515278101, + 0.01440343540161848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.34870266914367676 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890454575419426 + ], + "opposite_logprob_tokens": [ + -2.464045524597168 + ], + "logprob_reward": -0.0890454575419426, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890454575419426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3614792823791504, + 0.10961420834064484, + 0.056086912751197815, + 0.09736821055412292, + 0.07345425337553024, + 0.06553611159324646, + 0.0429752953350544, + 0.06382320821285248, + 0.06993455439805984, + 0.05972800403833389 + ], + "credit_map": [ + 0.020485104992985725, + 0.016985345631837845, + 0.01835009455680847, + 0.023629961535334587, + 0.02093648537993431, + 0.1202189028263092, + 0.01976599171757698, + 0.12110736966133118, + 0.01296680886298418, + 0.007756226696074009, + 0.008970063179731369, + 0.012641937471926212, + 0.006732757203280926, + 0.0230998694896698, + 0.011087380349636078, + 0.02635917067527771, + 0.00852458830922842, + 0.0054067587479949, + 0.006098486948758364, + 0.006798464339226484, + 0.004312812816351652, + 0.009057735092937946, + 0.00728909345343709, + 0.008598974905908108, + 0.01695111393928528, + 0.01213045697659254, + 0.010973740369081497, + 0.009222954511642456, + 0.007213293574750423, + 0.016800640150904655, + 0.00939319096505642, + 0.014682830311357975, + 0.005019015166908503, + 0.0032845286186784506, + 0.022491512820124626, + 0.006878409534692764, + 0.004848300013691187, + 0.02259635552763939, + 0.004534412175416946, + 0.0038017157930880785, + 0.005586313083767891, + 0.003654091851785779, + 0.019124045968055725, + 0.0069837067276239395, + 0.005244236905127764, + 0.012120334431529045, + 0.007463006302714348, + 0.005360380280762911, + 0.00448756106197834, + 0.003026198362931609, + 0.011711438186466694, + 0.0049763936549425125, + 0.002440756419673562, + 0.007140589412301779, + 0.00451632309705019, + 0.0046760388650000095, + 0.003981777001172304, + 0.003499315120279789, + 0.01578265242278576, + 0.005520871840417385, + 0.003016873262822628, + 0.009825876913964748, + 0.004214160609990358, + 0.01798168383538723, + 0.004743838682770729, + 0.006448141299188137, + 0.0047369785606861115, + 0.00591350207105279, + 0.006231657695025206, + 0.015707952901721, + 0.007292582653462887, + 0.018859898671507835, + 0.006631701253354549, + 0.006824626121670008, + 0.005390744656324387, + 0.005742256063967943, + 0.006217698100954294, + 0.010536770336329937, + 0.0059021771885454655, + 0.012482029385864735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0890454575419426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme macro close-up, captured in Annie Leibovitz's signature austere portrait style, the camera remains motionless, darting incredibly close to the face of a figure half-lying in bed. The figure has just cried, and a soft, directed studio light cast from above and to the side precisely captures the traces of tears still lingering at the corners of their eyes. The wet trails create sharp, soft highlights on their skin, causing a few eyelashes to cling together with a slight dampness. The lens is sharply focused on a slender, transparent medical plastic tube, which conforms softly to the contours of the skin. The end of the tube is gently inserted into one of the figure's nostrils, seemingly delivering a barely perceptible stream of air. The tube's walls appear subtly textured in the light. The extreme close-up reveals the figure's skin texture and the finest pores, while the facial muscles appear relaxed and weary after weeping. The background is a completely out-of-focus, soft white bed sheet, forming a pure and blurred block of color, which draws all visual focus to this core detail full of fragility and vitality. The whole picture conveys an atmosphere of profound intimacy, sympathy and human dignity in silence.", + "seed": 19659, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_6/index_6955_seed_19659_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2115548588335514, + "vlm_logprob": -0.2115548588335514, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.2115548588335514, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030315756797790527 + ], + "opposite_logprob_tokens": [ + -4.15531587600708 + ], + "logprob_reward": -0.030315756797790527, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.030315756797790527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2972715497016907, + 0.18778270483016968, + 0.04162178188562393, + 0.04845185577869415, + 0.05482262372970581, + 0.0373372919857502, + 0.04793372005224228, + 0.04992765560746193, + 0.1317753940820694, + 0.10307541489601135 + ], + "credit_map": [ + 0.03268216550350189, + 0.028124719858169556, + 0.037773799151182175, + 0.04013974592089653, + 0.02242589369416237, + 0.03309936448931694, + 0.058309752494096756, + 0.044716112315654755, + 0.021976012736558914, + 0.024603893980383873, + 0.019864827394485474, + 0.02198709174990654, + 0.014450042508542538, + 0.02720310352742672, + 0.0315517857670784, + 0.02614595554769039, + 0.005976981483399868, + 0.005439621862024069, + 0.003897314425557852, + 0.009205257520079613, + 0.002080270554870367, + 0.002609310671687126, + 0.005656112916767597, + 0.006756912916898727, + 0.0070244912058115005, + 0.008754624985158443, + 0.004463496617972851, + 0.007581639103591442, + 0.0032470035366714, + 0.005217652767896652, + 0.005875703413039446, + 0.006287246011197567, + 0.0068427082151174545, + 0.00686158612370491, + 0.006604279391467571, + 0.0063248141668736935, + 0.004274868406355381, + 0.006527030374854803, + 0.010274344123899937, + 0.007112991530448198, + 0.004656006582081318, + 0.003350990591570735, + 0.004038301296532154, + 0.006550543010234833, + 0.0028520456980913877, + 0.0033967476338148117, + 0.00703006936237216, + 0.005462589208036661, + 0.005844381172209978, + 0.005513114854693413, + 0.005181929562240839, + 0.006972004193812609, + 0.0035188726615160704, + 0.003857105039060116, + 0.010421572253108025, + 0.006624739151448011, + 0.006893977988511324, + 0.007344396319240332, + 0.007048321887850761, + 0.007993592880666256, + 0.003062100615352392, + 0.0039624618366360664, + 0.006638790480792522, + 0.006984017323702574, + 0.02603280358016491, + 0.01349274255335331, + 0.009715295396745205, + 0.014642483554780483, + 0.01039196364581585, + 0.017102666199207306, + 0.015873489901423454, + 0.02452394738793373, + 0.02757841907441616, + 0.011486387811601162, + 0.006492988206446171, + 0.011253302916884422, + 0.004213796928524971, + 0.0058935657143592834, + 0.007120051886886358, + 0.029036903753876686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.030315756797790527 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578379511833191 + ], + "opposite_logprob_tokens": [ + -0.6328379511833191 + ], + "logprob_reward": -0.7578379511833191, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578379511833191, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1727529764175415, + 0.05957931652665138, + 0.03375152125954628, + 0.03023502416908741, + 0.04924977198243141, + 0.04480985924601555, + 0.16084134578704834, + 0.2538980543613434, + 0.16895604133605957, + 0.025926029309630394 + ], + "credit_map": [ + 0.021714957430958748, + 0.03301185369491577, + 0.01635083183646202, + 0.030258776620030403, + 0.010365447029471397, + 0.016816336661577225, + 0.023701386526226997, + 0.020533397793769836, + 0.009155395440757275, + 0.007847405970096588, + 0.006162680219858885, + 0.008590714074671268, + 0.0036791348829865456, + 0.006257200613617897, + 0.00868566520512104, + 0.009201129898428917, + 0.0037097837775945663, + 0.004233444109559059, + 0.0035540752578526735, + 0.006275363732129335, + 0.0023816758766770363, + 0.0026877133641391993, + 0.006559872999787331, + 0.004349593538790941, + 0.0036857586819678545, + 0.0032720095477998257, + 0.0035871623549610376, + 0.005112631246447563, + 0.0021734824404120445, + 0.0031332385260611773, + 0.004823832307010889, + 0.004446910694241524, + 0.0068164849653840065, + 0.006795957218855619, + 0.005477185361087322, + 0.008743160404264927, + 0.0039605265483260155, + 0.005952820647507906, + 0.004815375432372093, + 0.006688268855214119, + 0.006173884030431509, + 0.005192379467189312, + 0.004160616081207991, + 0.006109550595283508, + 0.0043861232697963715, + 0.005259523168206215, + 0.007149299141019583, + 0.006378487218171358, + 0.03217822685837746, + 0.018433840945363045, + 0.012244259007275105, + 0.02476084791123867, + 0.010584533214569092, + 0.012106243520975113, + 0.014851267449557781, + 0.03568213805556297, + 0.06561856716871262, + 0.028207655996084213, + 0.015931256115436554, + 0.023655995726585388, + 0.012170538306236267, + 0.016745567321777344, + 0.027774717658758163, + 0.06379376351833344, + 0.039651624858379364, + 0.01951809599995613, + 0.013452872633934021, + 0.02425631321966648, + 0.00803715642541647, + 0.011140373535454273, + 0.012509280815720558, + 0.04039035364985466, + 0.004385903477668762, + 0.0038328799419105053, + 0.0028619635850191116, + 0.0033021830022335052, + 0.0018662101356312633, + 0.0027498274575918913, + 0.003111770609393716, + 0.0038152930792421103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.7578379511833191 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206582680344582 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206582680344582, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206582680344582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3408467173576355, + 0.09981297701597214, + 0.11671741306781769, + 0.05884891748428345, + 0.054038770496845245, + 0.06858081370592117, + 0.07454139739274979, + 0.061551399528980255, + 0.056060757488012314, + 0.06900086253881454 + ], + "credit_map": [ + 0.03277331590652466, + 0.03585956245660782, + 0.037483274936676025, + 0.05355536937713623, + 0.021360881626605988, + 0.030482547357678413, + 0.08767479658126831, + 0.0416569709777832, + 0.01183316670358181, + 0.009507150389254093, + 0.01064297091215849, + 0.01615893468260765, + 0.007645846344530582, + 0.011548210866749287, + 0.01904970407485962, + 0.01342698372900486, + 0.013372133485972881, + 0.012849885039031506, + 0.011772233992815018, + 0.01631481572985649, + 0.00859593041241169, + 0.013011499308049679, + 0.02601919136941433, + 0.014781713485717773, + 0.006421929690986872, + 0.008691745810210705, + 0.0057524945586919785, + 0.009946434758603573, + 0.003957298118621111, + 0.005334862042218447, + 0.01151222363114357, + 0.007231921423226595, + 0.006535917054861784, + 0.007520063780248165, + 0.0056206355802714825, + 0.007511560805141926, + 0.0042139822617173195, + 0.0047961147502064705, + 0.010515701025724411, + 0.007324791979044676, + 0.010383639484643936, + 0.012162205763161182, + 0.008325448259711266, + 0.007534221280366182, + 0.004572715610265732, + 0.007988372817635536, + 0.010333209298551083, + 0.007280993741005659, + 0.011894828639924526, + 0.008843815885484219, + 0.008883259259164333, + 0.010680393315851688, + 0.004894601181149483, + 0.005545878782868385, + 0.011290866881608963, + 0.012507759965956211, + 0.01636812649667263, + 0.007533443160355091, + 0.003688601776957512, + 0.006222020834684372, + 0.0028560813516378403, + 0.003914136905223131, + 0.004691367968916893, + 0.01627761498093605, + 0.0076326774433255196, + 0.009314515627920628, + 0.005610673222690821, + 0.009079382754862309, + 0.004657168406993151, + 0.004163443576544523, + 0.009138324297964573, + 0.006464564241468906, + 0.015392647124826908, + 0.006977011915296316, + 0.005128799006342888, + 0.008126013912260532, + 0.004339403007179499, + 0.005608500447124243, + 0.007085119374096394, + 0.016343362629413605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014206582680344582 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256171852350235 + ], + "opposite_logprob_tokens": [ + -1.6006171703338623 + ], + "logprob_reward": -0.2256171852350235, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256171852350235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26270538568496704, + 0.19521665573120117, + 0.056630950421094894, + 0.029487015679478645, + 0.058845050632953644, + 0.11604787409305573, + 0.03844066336750984, + 0.027945933863520622, + 0.13995127379894257, + 0.07472918927669525 + ], + "credit_map": [ + 0.030586978420615196, + 0.033436477184295654, + 0.019427167251706123, + 0.03756355494260788, + 0.022618472576141357, + 0.04677444323897362, + 0.03973741456866264, + 0.032560866326093674, + 0.021067144349217415, + 0.023588523268699646, + 0.01916331611573696, + 0.0274385754019022, + 0.014806394465267658, + 0.030546419322490692, + 0.03406519442796707, + 0.024541089311242104, + 0.007375981193035841, + 0.0077689094468951225, + 0.003700871719047427, + 0.008339667692780495, + 0.004546954296529293, + 0.00661278935149312, + 0.011388671584427357, + 0.006897104438394308, + 0.003653191728517413, + 0.003562428057193756, + 0.0028308643959462643, + 0.005063548218458891, + 0.002641355851665139, + 0.003957345150411129, + 0.0038029986899346113, + 0.003975283354520798, + 0.006980338599532843, + 0.00650074053555727, + 0.006660290993750095, + 0.008626767434179783, + 0.0045857676304876804, + 0.00623831432312727, + 0.012129013426601887, + 0.007123817689716816, + 0.019274383783340454, + 0.02526281028985977, + 0.011650976724922657, + 0.015141301788389683, + 0.007889037020504475, + 0.008417883887887001, + 0.01738278940320015, + 0.011028682813048363, + 0.00495054991915822, + 0.00502407643944025, + 0.0034750690683722496, + 0.005314316134899855, + 0.002237575361505151, + 0.0036900476552546024, + 0.008391511626541615, + 0.0053575169295072556, + 0.00313140288926661, + 0.00311402790248394, + 0.004250018857419491, + 0.004148968029767275, + 0.002314448356628418, + 0.002999338787049055, + 0.004551099147647619, + 0.003436630591750145, + 0.03484277054667473, + 0.014184549450874329, + 0.009396608918905258, + 0.017776325345039368, + 0.006544406060129404, + 0.010019158013164997, + 0.00916670635342598, + 0.038020748645067215, + 0.0205323938280344, + 0.009383995085954666, + 0.004412303678691387, + 0.007540888153016567, + 0.003223200561478734, + 0.004154813941568136, + 0.004305084235966206, + 0.02117650769650936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.2256171852350235 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro close-up of a figure's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera completely motionless throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a slender, transparent medical plastic tube visible inserted into the figure's nostril?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are visible wet trails of tears at the corners of the eyes with sharp highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a completely out-of-focus, soft white bed sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796818271279335 + ], + "opposite_logprob_tokens": [ + -3.529796838760376 + ], + "logprob_reward": -0.029796818271279335, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796818271279335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11208074539899826, + 0.09081139415502548, + 0.03842148929834366, + 0.09276725351810455, + 0.04508048668503761, + 0.3340229094028473, + 0.034436728805303574, + 0.08295067399740219, + 0.0710967555642128, + 0.09833155572414398 + ], + "credit_map": [ + 0.014585517346858978, + 0.013606897555291653, + 0.01579783670604229, + 0.024020198732614517, + 0.007108370307832956, + 0.012160987593233585, + 0.011150926351547241, + 0.01365000382065773, + 0.012824544683098793, + 0.01341885607689619, + 0.011979940347373486, + 0.012859305366873741, + 0.006381823681294918, + 0.010435459204018116, + 0.012780023738741875, + 0.010131443850696087, + 0.00443835835903883, + 0.004799771122634411, + 0.003966891206800938, + 0.007085578516125679, + 0.0026239664293825626, + 0.0036008041352033615, + 0.006773447152227163, + 0.005132673308253288, + 0.012805075384676456, + 0.019702453166246414, + 0.01074213907122612, + 0.016066882759332657, + 0.005828538443893194, + 0.005249895621091127, + 0.010809595696628094, + 0.011562677100300789, + 0.005768890492618084, + 0.007434612140059471, + 0.005200751591473818, + 0.006528452970087528, + 0.004181605763733387, + 0.006008680909872055, + 0.005400520283728838, + 0.0045569706708192825, + 0.05132034048438072, + 0.0603034533560276, + 0.05335523188114166, + 0.047995857894420624, + 0.020575981587171555, + 0.022002005949616432, + 0.040585823357105255, + 0.0378841906785965, + 0.007731068879365921, + 0.003740217536687851, + 0.0027289148420095444, + 0.004569848999381065, + 0.0024978802539408207, + 0.0020973668433725834, + 0.0025899705942720175, + 0.008481462486088276, + 0.02138342335820198, + 0.012184725143015385, + 0.005626437719911337, + 0.00784893799573183, + 0.002766428980976343, + 0.004367866087704897, + 0.008968247100710869, + 0.019804613664746284, + 0.017157895490527153, + 0.009882702492177486, + 0.005852638743817806, + 0.009147340431809425, + 0.0027849343605339527, + 0.004249870777130127, + 0.00515516335144639, + 0.016866208985447884, + 0.01583985425531864, + 0.012647103518247604, + 0.013343926519155502, + 0.013394366949796677, + 0.0073323617689311504, + 0.008284901268780231, + 0.010855657048523426, + 0.016633382067084312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.029796818271279335 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_8/index_tensor([8795, 8795])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_8/index_tensor([8795, 8795])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..50543daf2141dedb3e05a7266c3d8c584579de49 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_8/index_tensor([8795, 8795])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium close-up captures the meticulous process of food preparation in a well-equipped kitchen. The focal point is a heavy, polished stainless steel meat grinder, securely clamped to the edge of a light-colored granite countertop. From its front grinding disc, coarse, deep-red, freshly ground meat is steadily squeezed, curled, and accumulated into a growing mound, deposited in a deep, shiny stainless steel bowl directly below. A human hand (visible from the forearm) continuously and methodically uses a cylindrical wooden plunger to push the deep-red, marbled chunks of raw meat into the grinder's top hopper. Against the softly blurred background, the warm tones of the wooden cabinets and the brushed metal surfaces of a tall stainless steel refrigerator create a sense of domestic depth. Other kitchen items, including a ceramic knife block and a small potted herb plant, are neatly arranged on the counter, further enhancing the scene's authenticity. The entire scene is illuminated by soft, warm overhead lighting, casting gentle reflections on the metal surfaces, creating an atmosphere of clean, focused culinary preparation that is efficient, almost therapeutic.", + "seed": 1462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_8/index_8795_seed_1462_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04590437561273575, + "vlm_logprob": -0.04590437561273575, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04590437561273575, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14871220290660858 + ], + "opposite_logprob_tokens": [ + -2.523712158203125 + ], + "logprob_reward": -0.14871220290660858, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.14871220290660858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2744857966899872, + 0.10470398515462875, + 0.09217561781406403, + 0.1007845476269722, + 0.06829407811164856, + 0.08416683226823807, + 0.10158659517765045, + 0.09447004646062851, + 0.049351755529642105, + 0.029980778694152832 + ], + "credit_map": [ + 0.030163612216711044, + 0.02553207240998745, + 0.02952101081609726, + 0.031704459339380264, + 0.043298278003931046, + 0.035162899643182755, + 0.04826497659087181, + 0.030838482081890106, + 0.012741626240313053, + 0.008349613286554813, + 0.011665313504636288, + 0.014004469849169254, + 0.013789891265332699, + 0.014640268869698048, + 0.017622314393520355, + 0.011890488676726818, + 0.017136501148343086, + 0.022524137049913406, + 0.010655752383172512, + 0.00974116101861, + 0.008022448979318142, + 0.008901889435946941, + 0.0073099033907055855, + 0.007883827202022076, + 0.013654837384819984, + 0.009365053847432137, + 0.016864432021975517, + 0.010945088230073452, + 0.017810294404625893, + 0.01601536013185978, + 0.00752650573849678, + 0.008602972142398357, + 0.006042692810297012, + 0.005812606308609247, + 0.006881063804030418, + 0.007933998480439186, + 0.008476333692669868, + 0.01363371405750513, + 0.008940993808209896, + 0.010572671890258789, + 0.007994784973561764, + 0.010429461486637592, + 0.009678144007921219, + 0.0077747562900185585, + 0.017345430329442024, + 0.011902082711458206, + 0.008663836866617203, + 0.010378335602581501, + 0.013588964939117432, + 0.010084973648190498, + 0.01276881992816925, + 0.010400393977761269, + 0.011843311600387096, + 0.014376389794051647, + 0.017348311841487885, + 0.011175435036420822, + 0.0135251609608531, + 0.016149483621120453, + 0.009870012290775776, + 0.01083082053810358, + 0.01174674928188324, + 0.011402874253690243, + 0.011636670678853989, + 0.00930827297270298, + 0.008120560087263584, + 0.005540494341403246, + 0.007445600349456072, + 0.005254050716757774, + 0.005100569687783718, + 0.007564079016447067, + 0.005285241175442934, + 0.005041155964136124, + 0.006387760862708092, + 0.005564589519053698, + 0.00383907463401556, + 0.003082812065258622, + 0.0025250280741602182, + 0.003163898829370737, + 0.002171122469007969, + 0.0032464920077472925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.14871220290660858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029830025508999825 + ], + "opposite_logprob_tokens": [ + -3.529829978942871 + ], + "logprob_reward": -0.029830025508999825, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029830025508999825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2847076654434204, + 0.28719422221183777, + 0.07912945747375488, + 0.06788459420204163, + 0.028114158660173416, + 0.027692005038261414, + 0.041852690279483795, + 0.028969237580895424, + 0.05518873780965805, + 0.09926721453666687 + ], + "credit_map": [ + 0.030785508453845978, + 0.024280207231640816, + 0.030407480895519257, + 0.03790789097547531, + 0.04250510036945343, + 0.03514735400676727, + 0.03844521567225456, + 0.04522889479994774, + 0.026913562789559364, + 0.029226016253232956, + 0.03258882090449333, + 0.03203567862510681, + 0.03902915120124817, + 0.04257982224225998, + 0.049401555210351944, + 0.035419613122940063, + 0.008898918516933918, + 0.005814566742628813, + 0.009436076506972313, + 0.007874599657952785, + 0.010177461430430412, + 0.007590108551084995, + 0.014192587696015835, + 0.015145131386816502, + 0.007429543882608414, + 0.005450946744531393, + 0.007555071264505386, + 0.007441082037985325, + 0.01053107064217329, + 0.01285875216126442, + 0.006340597290545702, + 0.01027753483504057, + 0.0033321387600153685, + 0.002403638791292906, + 0.004649971146136522, + 0.003940447699278593, + 0.0029203849844634533, + 0.0032547451555728912, + 0.004538067150861025, + 0.0030747647397220135, + 0.0029261105228215456, + 0.002779599977657199, + 0.0034113158471882343, + 0.002512165578082204, + 0.005195952951908112, + 0.003812304697930813, + 0.0044188350439071655, + 0.0026357199531048536, + 0.00629494758322835, + 0.003205504035577178, + 0.006050200201570988, + 0.004098120145499706, + 0.004822122864425182, + 0.005180024076253176, + 0.007897044532001019, + 0.00430472893640399, + 0.004300293512642384, + 0.003842593403533101, + 0.0026958000380545855, + 0.0030712247826159, + 0.003470281371846795, + 0.0031205511186271906, + 0.004922614898532629, + 0.003545877756550908, + 0.00567473191767931, + 0.004744669422507286, + 0.0071609243750572205, + 0.005469296593219042, + 0.010370045900344849, + 0.008931223303079605, + 0.007579129654914141, + 0.005258717108517885, + 0.008286656811833382, + 0.008035656064748764, + 0.009523947723209858, + 0.009365660138428211, + 0.013036172837018967, + 0.023435791954398155, + 0.016828885301947594, + 0.010754446499049664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.029830025508999825 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029808389022946358 + ], + "opposite_logprob_tokens": [ + -3.529808282852173 + ], + "logprob_reward": -0.029808389022946358, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029808389022946358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47480127215385437, + 0.09361732006072998, + 0.05967990309000015, + 0.084007628262043, + 0.03832823410630226, + 0.022884663194417953, + 0.02867087349295616, + 0.029440002515912056, + 0.1347985863685608, + 0.03377140313386917 + ], + "credit_map": [ + 0.03992132097482681, + 0.04017827287316322, + 0.04317032918334007, + 0.04401174932718277, + 0.116465725004673, + 0.0506293885409832, + 0.09801006317138672, + 0.04241446778178215, + 0.008408940397202969, + 0.010319415479898453, + 0.008540480397641659, + 0.008676434867084026, + 0.010726850479841232, + 0.011554568074643612, + 0.020749805495142937, + 0.014640836976468563, + 0.0096695302054286, + 0.006128436885774136, + 0.007463530637323856, + 0.004211893770843744, + 0.007494902703911066, + 0.009449041448533535, + 0.007329905405640602, + 0.007932673208415508, + 0.0074026803486049175, + 0.007090478669852018, + 0.009228806011378765, + 0.007688777055591345, + 0.01907355710864067, + 0.01727059856057167, + 0.008586670272052288, + 0.007666064891964197, + 0.004993901588022709, + 0.003358361078426242, + 0.0055513824336230755, + 0.0033850008621811867, + 0.00425802543759346, + 0.006491362117230892, + 0.004203814547508955, + 0.006086392793804407, + 0.003163183806464076, + 0.003442147048190236, + 0.003270774381235242, + 0.002399904653429985, + 0.0022832623217254877, + 0.002842986723408103, + 0.0027294030878692865, + 0.0027530028019100428, + 0.0038762420881539583, + 0.004029319155961275, + 0.0028884101193398237, + 0.0023615877144038677, + 0.0032241339795291424, + 0.0033312267623841763, + 0.005682562943547964, + 0.0032773930579423904, + 0.005943950731307268, + 0.004582908470183611, + 0.0031075074803084135, + 0.0022331662476062775, + 0.0034984261728823185, + 0.0034900568425655365, + 0.0034743899013847113, + 0.0031096015591174364, + 0.01380996871739626, + 0.01248254906386137, + 0.01169012300670147, + 0.012450972571969032, + 0.026531049981713295, + 0.022115718573331833, + 0.02179029770195484, + 0.01392793282866478, + 0.0026764923240989447, + 0.0032173623330891132, + 0.003742503235116601, + 0.0035139473620802164, + 0.0043500084429979324, + 0.008196826092898846, + 0.004356001038104296, + 0.003718264400959015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.029808389022946358 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634377270936966 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634377270936966, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634377270936966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27155354619026184, + 0.15894703567028046, + 0.06754373013973236, + 0.09630829840898514, + 0.04757509008049965, + 0.1209561750292778, + 0.048734720796346664, + 0.04514284059405327, + 0.04836351424455643, + 0.09487511217594147 + ], + "credit_map": [ + 0.02968495525419712, + 0.03749380260705948, + 0.030688682571053505, + 0.029909081757068634, + 0.03482022508978844, + 0.039836347103118896, + 0.03862608224153519, + 0.03049435093998909, + 0.016583368182182312, + 0.019883276894688606, + 0.0165611170232296, + 0.014312239363789558, + 0.01897553913295269, + 0.02135608345270157, + 0.03016088157892227, + 0.021114518865942955, + 0.0074271997436881065, + 0.007461877074092627, + 0.00756376376375556, + 0.004818319343030453, + 0.01045166701078415, + 0.012442606501281261, + 0.009343923069536686, + 0.008034366182982922, + 0.009753298945724964, + 0.007435968611389399, + 0.010538875125348568, + 0.008586326614022255, + 0.021107463166117668, + 0.018889307975769043, + 0.01014702394604683, + 0.009850027039647102, + 0.006106115411967039, + 0.00472596799954772, + 0.007105136290192604, + 0.0051552834920585155, + 0.005646281875669956, + 0.006466571241617203, + 0.007893525995314121, + 0.0044762068428099155, + 0.012616807594895363, + 0.014110402204096317, + 0.014114053919911385, + 0.009023371152579784, + 0.026279686018824577, + 0.017423108220100403, + 0.01440711971372366, + 0.012981612235307693, + 0.005288135726004839, + 0.006002611014991999, + 0.0048453486524522305, + 0.004098082426935434, + 0.004459581803530455, + 0.0062697106041014194, + 0.012379290536046028, + 0.005391954910010099, + 0.005346694495528936, + 0.0050153774209320545, + 0.004307167138904333, + 0.0037316596135497093, + 0.005749065428972244, + 0.006344449240714312, + 0.009520920924842358, + 0.005127502139657736, + 0.006563378497958183, + 0.0038775261491537094, + 0.00536075234413147, + 0.0043234010227024555, + 0.008914905600249767, + 0.00755901588127017, + 0.007522370666265488, + 0.004242162220180035, + 0.008443759754300117, + 0.0077703846618533134, + 0.01073737908154726, + 0.008034723810851574, + 0.011990274302661419, + 0.021970193833112717, + 0.015916980803012848, + 0.010011408478021622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.008634377270936966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536883354187012 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012536883354187012, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536883354187012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43003612756729126, + 0.11363730579614639, + 0.07177181541919708, + 0.06847553700208664, + 0.04671081528067589, + 0.06133167818188667, + 0.059488434344530106, + 0.029863107949495316, + 0.05406433716416359, + 0.06462093442678452 + ], + "credit_map": [ + 0.035280194133520126, + 0.05203292518854141, + 0.05097988247871399, + 0.05355387181043625, + 0.09356652200222015, + 0.045747872442007065, + 0.05940233916044235, + 0.039472464472055435, + 0.012055552564561367, + 0.01766437664628029, + 0.01103078480809927, + 0.011808519251644611, + 0.014265439473092556, + 0.015817467123270035, + 0.015141716226935387, + 0.015853436663746834, + 0.011937269940972328, + 0.006252995226532221, + 0.011937812902033329, + 0.010552278719842434, + 0.0062957387417554855, + 0.0075584957376122475, + 0.008240967057645321, + 0.008996251970529556, + 0.00664945226162672, + 0.00952361710369587, + 0.00678621931001544, + 0.006470871157944202, + 0.01178707741200924, + 0.013786613009870052, + 0.008006906136870384, + 0.005464770365506411, + 0.004810857120901346, + 0.004087102133780718, + 0.005914945155382156, + 0.004418476484715939, + 0.006256601307541132, + 0.00889676995575428, + 0.006651208270341158, + 0.0056748525239527225, + 0.006470056250691414, + 0.006878090091049671, + 0.008574424311518669, + 0.006153222173452377, + 0.012904899194836617, + 0.008306987583637238, + 0.005933476146310568, + 0.006110518705099821, + 0.00820851232856512, + 0.006122164893895388, + 0.00798011850565672, + 0.006736660376191139, + 0.006334883160889149, + 0.008358443155884743, + 0.007792361080646515, + 0.007955287583172321, + 0.004316984675824642, + 0.0031652774196118116, + 0.0037504227366298437, + 0.004758930765092373, + 0.003877876326441765, + 0.003948013763874769, + 0.003011806169524789, + 0.0030337939970195293, + 0.005357191897928715, + 0.006098236422985792, + 0.0054313442669808865, + 0.00473420973867178, + 0.009054227732121944, + 0.00897330790758133, + 0.008693888783454895, + 0.005721923895180225, + 0.00776176992803812, + 0.009867502376437187, + 0.006491188891232014, + 0.007423996925354004, + 0.007751488126814365, + 0.011345414444804192, + 0.00787363201379776, + 0.006105934269726276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.012536883354187012 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures the meticulous process of food preparation in a well-equipped kitchen. The focal point is a heavy, polished stainless steel meat grinder, securely clamped to the edge of a light-colored granite countertop. From its front grinding disc, coarse, deep-red, freshly ground meat is steadily squeezed, curled, and accumulated into a growing mound, deposited in a deep, shiny stainless steel bowl directly below. A human hand (visible from the forearm) continuously and methodically uses a cylindrical wooden plunger to push the deep-red, marbled chunks of raw meat into the grinder's top hopper. Against the softly blurred background, the warm tones of the wooden cabinets and the brushed metal surfaces of a tall stainless steel refrigerator create a sense of domestic depth. Other kitchen items, including a ceramic knife block and a small potted herb plant, are neatly arranged on the counter, further enhancing the scene's authenticity. The entire scene is illuminated by soft, warm overhead lighting, casting gentle reflections on the metal surfaces, creating an atmosphere of clean, focused culinary preparation that is efficient, almost therapeutic.", + "seed": 1463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_8/index_8795_seed_1463_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.032574193086475134, + "vlm_logprob": -0.032574193086475134, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.032574193086475134, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07247284054756165 + ], + "opposite_logprob_tokens": [ + -3.1974728107452393 + ], + "logprob_reward": -0.07247284054756165, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.07247284054756165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2205430120229721, + 0.09858311712741852, + 0.07731586694717407, + 0.09667602926492691, + 0.13495366275310516, + 0.0809437558054924, + 0.09339089691638947, + 0.06801566481590271, + 0.08765647560358047, + 0.04192154482007027 + ], + "credit_map": [ + 0.027084728702902794, + 0.025003572925925255, + 0.027402730658650398, + 0.023408913984894753, + 0.0350792296230793, + 0.03196266293525696, + 0.03103504329919815, + 0.01956612803041935, + 0.009851210750639439, + 0.01754382997751236, + 0.005850310903042555, + 0.009533144533634186, + 0.0111161470413208, + 0.02451261878013611, + 0.01077526155859232, + 0.009400598704814911, + 0.012190372683107853, + 0.010316351428627968, + 0.007270685397088528, + 0.008910068310797215, + 0.008708257228136063, + 0.010030186735093594, + 0.010335243307054043, + 0.009554700925946236, + 0.011890213936567307, + 0.01574489288032055, + 0.009691204875707626, + 0.010734004899859428, + 0.009519313462078571, + 0.021317357197403908, + 0.0072027407586574554, + 0.010576304979622364, + 0.019358422607183456, + 0.019642630591988564, + 0.016805456951260567, + 0.014973209239542484, + 0.01234748400747776, + 0.02312784269452095, + 0.014142962172627449, + 0.01455564983189106, + 0.01382092759013176, + 0.012623021379113197, + 0.006429294589906931, + 0.006731972564011812, + 0.014472773298621178, + 0.009777741506695747, + 0.008968068286776543, + 0.008119954727590084, + 0.018911296501755714, + 0.011685916222631931, + 0.005210176110267639, + 0.008563409559428692, + 0.009233602322638035, + 0.024062199518084526, + 0.00846806075423956, + 0.007256240118294954, + 0.008828854188323021, + 0.011410519480705261, + 0.006963938474655151, + 0.007489497773349285, + 0.008227330632507801, + 0.00859126728028059, + 0.007573149167001247, + 0.0089311134070158, + 0.013383707031607628, + 0.019018402323126793, + 0.005765977315604687, + 0.009082711301743984, + 0.008240394294261932, + 0.019969096407294273, + 0.006265740375965834, + 0.005930446088314056, + 0.006901373155415058, + 0.005871207918971777, + 0.0031963856890797615, + 0.004713832400739193, + 0.009332552552223206, + 0.004479621537029743, + 0.003996343817561865, + 0.0034302298445254564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.07247284054756165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076476410031318665 + ], + "opposite_logprob_tokens": [ + -4.882647514343262 + ], + "logprob_reward": -0.0076476410031318665, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076476410031318665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41523244976997375, + 0.10405987501144409, + 0.07808981835842133, + 0.03280705213546753, + 0.028495507314801216, + 0.04658116027712822, + 0.06875234842300415, + 0.061266373842954636, + 0.1276390105485916, + 0.03707636520266533 + ], + "credit_map": [ + 0.03157171607017517, + 0.12178124487400055, + 0.023186225444078445, + 0.022737732157111168, + 0.026053590700030327, + 0.14787374436855316, + 0.02077079750597477, + 0.02125745452940464, + 0.010652094148099422, + 0.017734356224536896, + 0.011711345985531807, + 0.009406517259776592, + 0.012329420074820518, + 0.021627994254231453, + 0.010311681777238846, + 0.010286468081176281, + 0.012848308309912682, + 0.012733176350593567, + 0.007287835702300072, + 0.006996685639023781, + 0.01155812107026577, + 0.007665195502340794, + 0.010536782443523407, + 0.008463720791041851, + 0.00467852083966136, + 0.003867002669721842, + 0.0035920392256230116, + 0.003981750924140215, + 0.004206705838441849, + 0.004347556736320257, + 0.004040686879307032, + 0.004092791583389044, + 0.002849578158929944, + 0.0033377627842128277, + 0.0027718115597963333, + 0.0036861510016024113, + 0.003843071637675166, + 0.003946039360016584, + 0.0045469398610293865, + 0.0035141557455062866, + 0.006758558563888073, + 0.00871850736439228, + 0.004424776881933212, + 0.005170564632862806, + 0.006885684095323086, + 0.006035405211150646, + 0.004475930705666542, + 0.004111741203814745, + 0.00899521168321371, + 0.011904468759894371, + 0.002302047796547413, + 0.005604717880487442, + 0.003914763685315847, + 0.027019713073968887, + 0.0033824429847300053, + 0.005628988146781921, + 0.007282469887286425, + 0.00858084112405777, + 0.006350778508931398, + 0.006952682044357061, + 0.010762221179902554, + 0.009277021512389183, + 0.005531736649572849, + 0.006528633181005716, + 0.012130961753427982, + 0.016055863350629807, + 0.008526649326086044, + 0.01252709235996008, + 0.021627064794301987, + 0.016168255358934402, + 0.016937334090471268, + 0.023665793240070343, + 0.0027693298179656267, + 0.005688738543540239, + 0.00219193403609097, + 0.0032942851539701223, + 0.004753172397613525, + 0.007713472470641136, + 0.005579903721809387, + 0.005085533484816551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0076476410031318665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002759724855423 + ], + "opposite_logprob_tokens": [ + -2.6950275897979736 + ], + "logprob_reward": -0.07002759724855423, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002759724855423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5089190602302551, + 0.0755019411444664, + 0.11150749027729034, + 0.026008008047938347, + 0.03527398034930229, + 0.03839022293686867, + 0.024783818051218987, + 0.05136042460799217, + 0.1054474413394928, + 0.022807631641626358 + ], + "credit_map": [ + 0.03832438588142395, + 0.1462140679359436, + 0.02530769258737564, + 0.030937105417251587, + 0.033210694789886475, + 0.18430861830711365, + 0.024889163672924042, + 0.025727292522788048, + 0.009130686521530151, + 0.010839905589818954, + 0.00823537539690733, + 0.007156771142035723, + 0.011348751373589039, + 0.011351707391440868, + 0.008753600530326366, + 0.008685141801834106, + 0.02081124484539032, + 0.017829062417149544, + 0.011117234826087952, + 0.011487002484500408, + 0.015368327498435974, + 0.015517094172537327, + 0.010559012182056904, + 0.008818510919809341, + 0.0027418560348451138, + 0.004162750206887722, + 0.002499002031981945, + 0.003110517980530858, + 0.0037880653981119394, + 0.002541354624554515, + 0.004089789930731058, + 0.003074670908972621, + 0.0028900124598294497, + 0.004438497591763735, + 0.002830671612173319, + 0.004116661846637726, + 0.004828205797821283, + 0.0033491405192762613, + 0.008450995199382305, + 0.0043697962537407875, + 0.006218498107045889, + 0.005394537001848221, + 0.005372797604650259, + 0.004610112402588129, + 0.005117079243063927, + 0.004618585109710693, + 0.003863941179588437, + 0.0031946736853569746, + 0.004524092189967632, + 0.0038669363129884005, + 0.0020759187173098326, + 0.002779990201815963, + 0.002432052744552493, + 0.002488785656169057, + 0.003806684399023652, + 0.0028093578293919563, + 0.009146097116172314, + 0.008007986471056938, + 0.0056442636996507645, + 0.005415759049355984, + 0.007352878339588642, + 0.007247889414429665, + 0.004278354346752167, + 0.004267198033630848, + 0.020103394985198975, + 0.019423650577664375, + 0.00689523434266448, + 0.010524640791118145, + 0.012826829217374325, + 0.02300587110221386, + 0.006247968878597021, + 0.006419854704290628, + 0.0037980261258780956, + 0.003608907340094447, + 0.002258045831695199, + 0.002556947525590658, + 0.0028535923920571804, + 0.004463295917958021, + 0.0015743252588436007, + 0.0016944929957389832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.07002759724855423 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091464448720217 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.004091464448720217, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091464448720217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5570841431617737, + 0.05304863303899765, + 0.03508087247610092, + 0.07285509258508682, + 0.059025321155786514, + 0.04576994106173515, + 0.052368663251399994, + 0.031094400212168694, + 0.07156269252300262, + 0.02211027219891548 + ], + "credit_map": [ + 0.02533695288002491, + 0.18001386523246765, + 0.014263749122619629, + 0.03141622245311737, + 0.03413654491305351, + 0.23440463840961456, + 0.016946667805314064, + 0.020565466955304146, + 0.004649807699024677, + 0.005170084070414305, + 0.006184552796185017, + 0.007539194077253342, + 0.009509103372693062, + 0.006934622768312693, + 0.006280058529227972, + 0.006781209260225296, + 0.006539963185787201, + 0.004602930974215269, + 0.0036241149064153433, + 0.0035994192585349083, + 0.0044907680712640285, + 0.0044584013521671295, + 0.00416944082826376, + 0.0035958343651145697, + 0.006178774870932102, + 0.008017579093575478, + 0.007119409739971161, + 0.010300642810761929, + 0.009316915646195412, + 0.008377301506698132, + 0.01252633798867464, + 0.011018131859600544, + 0.005424981005489826, + 0.007135970052331686, + 0.005696744192391634, + 0.0072243777103722095, + 0.008672553114593029, + 0.007370042148977518, + 0.0102523909881711, + 0.007248266134411097, + 0.0052223848178982735, + 0.005417523439973593, + 0.0030697849579155445, + 0.004463670775294304, + 0.011028016917407513, + 0.007205979432910681, + 0.004736872389912605, + 0.004625710193067789, + 0.005253625102341175, + 0.01109378319233656, + 0.002242987509816885, + 0.003351191058754921, + 0.004541946575045586, + 0.013730872422456741, + 0.00524865323677659, + 0.0069055999629199505, + 0.0031645288690924644, + 0.005669200327247381, + 0.0033116000704467297, + 0.0034419922158122063, + 0.003812938230112195, + 0.004618978593498468, + 0.004089805297553539, + 0.0029853591695427895, + 0.010735665448009968, + 0.015200513415038586, + 0.006368733011186123, + 0.006793217733502388, + 0.008650161325931549, + 0.013754191808402538, + 0.005037750583142042, + 0.0050224605947732925, + 0.00274220434948802, + 0.0030105102341622114, + 0.001802323735319078, + 0.002557396190240979, + 0.002830488607287407, + 0.004093542695045471, + 0.002851287368685007, + 0.0022225198335945606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.004091464448720217 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008631422184407711 + ], + "opposite_logprob_tokens": [ + -4.758631229400635 + ], + "logprob_reward": -0.008631422184407711, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008631422184407711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3458369970321655, + 0.07862307876348495, + 0.08734743297100067, + 0.05356787517666817, + 0.08764491975307465, + 0.05696823447942734, + 0.1389162391424179, + 0.027513669803738594, + 0.08552791178226471, + 0.03805368393659592 + ], + "credit_map": [ + 0.023482495918869972, + 0.09847661107778549, + 0.012788957916200161, + 0.022120773792266846, + 0.021923266351222992, + 0.12260626256465912, + 0.0266988817602396, + 0.01773974299430847, + 0.01040471252053976, + 0.010894149541854858, + 0.005685270298272371, + 0.009225184097886086, + 0.011993414722383022, + 0.009705470874905586, + 0.011055034585297108, + 0.009659829549491405, + 0.01918349787592888, + 0.01234580110758543, + 0.007420709356665611, + 0.010018766857683659, + 0.009962632320821285, + 0.010413243435323238, + 0.0105682872235775, + 0.007434491068124771, + 0.00617626216262579, + 0.008127073757350445, + 0.005866692867130041, + 0.00783410482108593, + 0.005965011194348335, + 0.006704818457365036, + 0.008245008066296577, + 0.004648903850466013, + 0.01907212659716606, + 0.014031430706381798, + 0.008881809189915657, + 0.013959914445877075, + 0.007951595820486546, + 0.009070801548659801, + 0.00798920076340437, + 0.006688043475151062, + 0.009568811394274235, + 0.009315336123108864, + 0.00526020210236311, + 0.00677117332816124, + 0.010196400806307793, + 0.006450337823480368, + 0.004684921354055405, + 0.004721055272966623, + 0.013065447099506855, + 0.032549962401390076, + 0.008993938565254211, + 0.011810485273599625, + 0.007911937311291695, + 0.05012726038694382, + 0.007407300174236298, + 0.007049896754324436, + 0.0035412677098065615, + 0.005181971937417984, + 0.002601419575512409, + 0.002770527033135295, + 0.003058522241190076, + 0.00492751644924283, + 0.003358352929353714, + 0.002074091462418437, + 0.015840811654925346, + 0.02057083696126938, + 0.010428357869386673, + 0.010298679582774639, + 0.0074223303236067295, + 0.010017551481723785, + 0.005891803652048111, + 0.005057538393884897, + 0.006096639670431614, + 0.0054777017794549465, + 0.0056058913469314575, + 0.00442738551646471, + 0.003429945558309555, + 0.005467903334647417, + 0.0033535438124090433, + 0.004194672219455242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.008631422184407711 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures the meticulous process of food preparation in a well-equipped kitchen. The focal point is a heavy, polished stainless steel meat grinder, securely clamped to the edge of a light-colored granite countertop. From its front grinding disc, coarse, deep-red, freshly ground meat is steadily squeezed, curled, and accumulated into a growing mound, deposited in a deep, shiny stainless steel bowl directly below. A human hand (visible from the forearm) continuously and methodically uses a cylindrical wooden plunger to push the deep-red, marbled chunks of raw meat into the grinder's top hopper. Against the softly blurred background, the warm tones of the wooden cabinets and the brushed metal surfaces of a tall stainless steel refrigerator create a sense of domestic depth. Other kitchen items, including a ceramic knife block and a small potted herb plant, are neatly arranged on the counter, further enhancing the scene's authenticity. The entire scene is illuminated by soft, warm overhead lighting, casting gentle reflections on the metal surfaces, creating an atmosphere of clean, focused culinary preparation that is efficient, almost therapeutic.", + "seed": 1464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_8/index_8795_seed_1464_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027583714667707683, + "vlm_logprob": -0.027583714667707683, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.027583714667707683, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10285815596580505 + ], + "opposite_logprob_tokens": [ + -2.727858066558838 + ], + "logprob_reward": -0.10285815596580505, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.10285815596580505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16993914544582367, + 0.3669493794441223, + 0.06721960008144379, + 0.037938192486763, + 0.03858925774693489, + 0.04541390389204025, + 0.12872318923473358, + 0.03077201545238495, + 0.07542110234498978, + 0.039034225046634674 + ], + "credit_map": [ + 0.028991542756557465, + 0.025280063971877098, + 0.022977204993367195, + 0.016943100839853287, + 0.02281014248728752, + 0.021280966699123383, + 0.01304610911756754, + 0.01861000806093216, + 0.07441352307796478, + 0.0873161256313324, + 0.07543826103210449, + 0.03444081172347069, + 0.025725672021508217, + 0.03494476526975632, + 0.01730349101126194, + 0.017366737127304077, + 0.014885550364851952, + 0.010793015360832214, + 0.012022546492516994, + 0.0061113592237234116, + 0.005348240025341511, + 0.00947803445160389, + 0.0035363011993467808, + 0.005044554825872183, + 0.009523719549179077, + 0.003956477623432875, + 0.005578736308962107, + 0.005343969911336899, + 0.003058306174352765, + 0.004703382030129433, + 0.00267603131942451, + 0.0030975721310824156, + 0.007011885289102793, + 0.005649879574775696, + 0.005892578512430191, + 0.0034207566641271114, + 0.0042121754959225655, + 0.005689697805792093, + 0.003751503536477685, + 0.0029607792384922504, + 0.01003262959420681, + 0.00852042343467474, + 0.007173285353928804, + 0.004972726572304964, + 0.005415863357484341, + 0.002244761446490884, + 0.004447842016816139, + 0.0026063735131174326, + 0.03725915029644966, + 0.011918635107576847, + 0.02326706424355507, + 0.01049118023365736, + 0.011655004695057869, + 0.014418717473745346, + 0.012325634248554707, + 0.007387807592749596, + 0.006806539837270975, + 0.0030173107516020536, + 0.004840206354856491, + 0.003197271842509508, + 0.002486124401912093, + 0.0036469053011387587, + 0.0030513033270835876, + 0.003726353170350194, + 0.011773716658353806, + 0.007835011929273605, + 0.010821345262229443, + 0.008596672676503658, + 0.009181169793009758, + 0.007612028159201145, + 0.011319277808070183, + 0.008281879127025604, + 0.00584327382966876, + 0.004881937522441149, + 0.0062033855356276035, + 0.008399673737585545, + 0.0027954585384577513, + 0.0036107581108808517, + 0.003972391132265329, + 0.0033273487351834774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.10285815596580505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011099852621555328 + ], + "opposite_logprob_tokens": [ + -4.511099815368652 + ], + "logprob_reward": -0.011099852621555328, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011099852621555328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17442071437835693, + 0.18916475772857666, + 0.08996816724538803, + 0.07760629802942276, + 0.03956502303481102, + 0.030885109677910805, + 0.051892735064029694, + 0.04044399783015251, + 0.23624473810195923, + 0.0698084607720375 + ], + "credit_map": [ + 0.03726784884929657, + 0.01949976198375225, + 0.028134360909461975, + 0.01603902131319046, + 0.016110554337501526, + 0.022151779383420944, + 0.018832998350262642, + 0.016384392976760864, + 0.03302937000989914, + 0.030769256874918938, + 0.032238028943538666, + 0.01753273233771324, + 0.015990784391760826, + 0.02968447096645832, + 0.013265362940728664, + 0.01665475405752659, + 0.023937124758958817, + 0.010573772713541985, + 0.015094755217432976, + 0.008032700978219509, + 0.00805036723613739, + 0.00834444910287857, + 0.007536621298640966, + 0.008398381061851978, + 0.018778756260871887, + 0.00874759815633297, + 0.013448256999254227, + 0.008071117103099823, + 0.006658285390585661, + 0.010817274451255798, + 0.00495563168078661, + 0.006129380315542221, + 0.006040530279278755, + 0.004294335376471281, + 0.005805076565593481, + 0.004211084451526403, + 0.005681994371116161, + 0.0031669228337705135, + 0.005455744452774525, + 0.004909335635602474, + 0.006039061583578587, + 0.0042569381184875965, + 0.004858244210481644, + 0.003015425056219101, + 0.003583478508517146, + 0.002380243968218565, + 0.0036667953245341778, + 0.003084922442212701, + 0.013456104323267937, + 0.005497889127582312, + 0.00878954492509365, + 0.004718700423836708, + 0.004125526174902916, + 0.008988485671579838, + 0.0029836168978363276, + 0.0033328651916235685, + 0.006673410534858704, + 0.004685158375650644, + 0.0058008949272334576, + 0.00499814935028553, + 0.0038121487013995647, + 0.00298295053653419, + 0.005859808996319771, + 0.005631474778056145, + 0.0309141818434, + 0.02411034144461155, + 0.032472770661115646, + 0.026295140385627747, + 0.029709191992878914, + 0.025550339370965958, + 0.04172062128782272, + 0.025472158566117287, + 0.008572899736464024, + 0.006346080452203751, + 0.008837942034006119, + 0.008563707582652569, + 0.009113878011703491, + 0.005893161054700613, + 0.013232193887233734, + 0.009248602204024792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.011099852621555328 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638768292963505 + ], + "opposite_logprob_tokens": [ + -4.882638931274414 + ], + "logprob_reward": -0.007638768292963505, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638768292963505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07230255752801895, + 0.752077579498291, + 0.0279662124812603, + 0.018481701612472534, + 0.014375205151736736, + 0.013741584494709969, + 0.024719243869185448, + 0.011967793107032776, + 0.032150376588106155, + 0.03221771866083145 + ], + "credit_map": [ + 0.015689503401517868, + 0.009008529596030712, + 0.010964388027787209, + 0.008372310549020767, + 0.008400666527450085, + 0.007340810727328062, + 0.005309620872139931, + 0.007216728758066893, + 0.13311007618904114, + 0.2000798135995865, + 0.1458960473537445, + 0.08776500076055527, + 0.049557678401470184, + 0.06742604076862335, + 0.03005305863916874, + 0.038189906626939774, + 0.00812328141182661, + 0.004156995099037886, + 0.005608152132481337, + 0.002855447120964527, + 0.002174043795093894, + 0.0015118139563128352, + 0.0013908438850194216, + 0.002145633567124605, + 0.0054892199113965034, + 0.0019565275870263577, + 0.003662779927253723, + 0.0018415747908875346, + 0.0017142272554337978, + 0.0014904610579833388, + 0.0010424465872347355, + 0.0012844636803492904, + 0.0031447866931557655, + 0.0013196720974519849, + 0.0022300230339169502, + 0.0016105129616335034, + 0.002109871245920658, + 0.0018299783114343882, + 0.001037022564560175, + 0.0010933371959254146, + 0.0026338924653828144, + 0.001342733041383326, + 0.0021836727391928434, + 0.0014365060487762094, + 0.0014754480216652155, + 0.002169863088056445, + 0.001028304803185165, + 0.0014711637049913406, + 0.003499785903841257, + 0.002658219076693058, + 0.003398799104616046, + 0.0022002989426255226, + 0.003718605265021324, + 0.0025434375274926424, + 0.0038878382183611393, + 0.0028122602961957455, + 0.0020754006691277027, + 0.0009994065621867776, + 0.0017323370557278395, + 0.0016995284240692854, + 0.0012148629175499082, + 0.0009566599619574845, + 0.00149892910849303, + 0.0017906688153743744, + 0.004543193615972996, + 0.0032152857165783644, + 0.004353655967861414, + 0.004330577794462442, + 0.004308563657104969, + 0.0038765561766922474, + 0.004514716565608978, + 0.003007827326655388, + 0.00422054436057806, + 0.002468266524374485, + 0.0043236310593783855, + 0.004238676279783249, + 0.004193570464849472, + 0.004528023302555084, + 0.005000409670174122, + 0.0032445958349853754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.007638768292963505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052503556944429874 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.0052503556944429874, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052503556944429874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13067175447940826, + 0.458327054977417, + 0.046411190181970596, + 0.04049088433384895, + 0.052530232816934586, + 0.04724254831671715, + 0.047661323100328445, + 0.03582113981246948, + 0.10805381089448929, + 0.032790083438158035 + ], + "credit_map": [ + 0.02754734270274639, + 0.01404639147222042, + 0.020650163292884827, + 0.012547600083053112, + 0.012837007641792297, + 0.015067514032125473, + 0.016055652871727943, + 0.011920087039470673, + 0.08925584703683853, + 0.10531680285930634, + 0.08542797714471817, + 0.050575725734233856, + 0.03303021937608719, + 0.04299083724617958, + 0.02632586844265461, + 0.025403771549463272, + 0.00974531564861536, + 0.009037706069648266, + 0.007353477645665407, + 0.004509337712079287, + 0.004268983844667673, + 0.003563258098438382, + 0.003821304300799966, + 0.004111809190362692, + 0.010564681142568588, + 0.004775128792971373, + 0.006923234090209007, + 0.004190992098301649, + 0.0031305800657719374, + 0.004545049741864204, + 0.0033253992442041636, + 0.0030358198564499617, + 0.008763913996517658, + 0.005275501403957605, + 0.0076426975429058075, + 0.00506216986104846, + 0.008358560502529144, + 0.0052067553624510765, + 0.007509343326091766, + 0.004711295012384653, + 0.009847975336015224, + 0.007725267205387354, + 0.007463171612471342, + 0.006182509940117598, + 0.0054901340045034885, + 0.0034156355541199446, + 0.004117338918149471, + 0.003000513883307576, + 0.007492843549698591, + 0.005675503518432379, + 0.006827984005212784, + 0.004527985118329525, + 0.006610818672925234, + 0.007691917009651661, + 0.004766436293721199, + 0.0040678344666957855, + 0.0057718404568731785, + 0.004526985343545675, + 0.005392656661570072, + 0.0039678397588431835, + 0.004312444012612104, + 0.0036476159002631903, + 0.004986281041055918, + 0.0032154759392142296, + 0.014192274771630764, + 0.014149028807878494, + 0.014635236002504826, + 0.012845311313867569, + 0.013399752788245678, + 0.010788803920149803, + 0.016987862065434456, + 0.01105553936213255, + 0.006416972726583481, + 0.0026828411500900984, + 0.004933751188218594, + 0.003464930923655629, + 0.004110932815819979, + 0.0029793591238558292, + 0.0050889248959720135, + 0.0031123696826398373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.0052503556944429874 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071440763771534 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071440763771534, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071440763771534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20390842854976654, + 0.2385157197713852, + 0.08105513453483582, + 0.052270304411649704, + 0.06729704141616821, + 0.044179871678352356, + 0.040175095200538635, + 0.039497680962085724, + 0.1471671313047409, + 0.08593357354402542 + ], + "credit_map": [ + 0.03995915502309799, + 0.023313181474804878, + 0.03182932361960411, + 0.01866697520017624, + 0.021724915131926537, + 0.04015662148594856, + 0.01329848077148199, + 0.01495975349098444, + 0.053298983722925186, + 0.05321497842669487, + 0.044614799320697784, + 0.024170683696866035, + 0.015048940666019917, + 0.02482978254556656, + 0.010253999382257462, + 0.013083558529615402, + 0.021277932450175285, + 0.011131608858704567, + 0.013745247386395931, + 0.00801459327340126, + 0.006029005628079176, + 0.008967563509941101, + 0.005046821665018797, + 0.006842359434813261, + 0.008943114429712296, + 0.00615900382399559, + 0.007037421688437462, + 0.0053042517974972725, + 0.006254552397876978, + 0.010060915723443031, + 0.0038614501245319843, + 0.004649594891816378, + 0.017976656556129456, + 0.007600920274853706, + 0.012922555208206177, + 0.006458422169089317, + 0.005967691075056791, + 0.0070669520646333694, + 0.004620411433279514, + 0.004683435894548893, + 0.008998469449579716, + 0.008043065667152405, + 0.006714545655995607, + 0.0034617690835148096, + 0.00499282730743289, + 0.004930652212351561, + 0.003936759661883116, + 0.003101782174780965, + 0.007327716331928968, + 0.004367027431726456, + 0.006330844014883041, + 0.004227877128869295, + 0.00325020938180387, + 0.004947233945131302, + 0.00631836149841547, + 0.003405824536457658, + 0.006409474648535252, + 0.004985722713172436, + 0.00561874732375145, + 0.003931513521820307, + 0.004358721897006035, + 0.0060349865816533566, + 0.004469351377338171, + 0.0036891622003167868, + 0.025287242606282234, + 0.016797533258795738, + 0.021987326443195343, + 0.017979439347982407, + 0.014637813903391361, + 0.02113332971930504, + 0.01743522845208645, + 0.011909203603863716, + 0.01429841574281454, + 0.007218217942863703, + 0.013098903931677341, + 0.007833637297153473, + 0.01001947931945324, + 0.018432896584272385, + 0.008583096787333488, + 0.006448929198086262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.011071440763771534 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up captures the meticulous process of food preparation in a well-equipped kitchen. The focal point is a heavy, polished stainless steel meat grinder, securely clamped to the edge of a light-colored granite countertop. From its front grinding disc, coarse, deep-red, freshly ground meat is steadily squeezed, curled, and accumulated into a growing mound, deposited in a deep, shiny stainless steel bowl directly below. A human hand (visible from the forearm) continuously and methodically uses a cylindrical wooden plunger to push the deep-red, marbled chunks of raw meat into the grinder's top hopper. Against the softly blurred background, the warm tones of the wooden cabinets and the brushed metal surfaces of a tall stainless steel refrigerator create a sense of domestic depth. Other kitchen items, including a ceramic knife block and a small potted herb plant, are neatly arranged on the counter, further enhancing the scene's authenticity. The entire scene is illuminated by soft, warm overhead lighting, casting gentle reflections on the metal surfaces, creating an atmosphere of clean, focused culinary preparation that is efficient, almost therapeutic.", + "seed": 1465, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000099/group_8/index_8795_seed_1465_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02743741199374199, + "vlm_logprob": -0.02743741199374199, + "vlm_yes_no_margin": 4.124999952316284, + "vlm_token_credit": -0.02743741199374199, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08775977045297623 + ], + "opposite_logprob_tokens": [ + -2.8377597332000732 + ], + "logprob_reward": -0.08775977045297623, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.08775977045297623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15192674100399017, + 0.14455556869506836, + 0.060597848147153854, + 0.1611797958612442, + 0.14002041518688202, + 0.036843471229076385, + 0.06995369493961334, + 0.13945826888084412, + 0.06520482152700424, + 0.030259346589446068 + ], + "credit_map": [ + 0.021786075085401535, + 0.031293705105781555, + 0.013480440713465214, + 0.017660845071077347, + 0.01959189958870411, + 0.013962794095277786, + 0.016810405999422073, + 0.01734057255089283, + 0.028643658384680748, + 0.030963830649852753, + 0.017673902213573456, + 0.014421040192246437, + 0.01797650009393692, + 0.011270269751548767, + 0.01229797676205635, + 0.011308390647172928, + 0.007269901689141989, + 0.008872242644429207, + 0.00810724776238203, + 0.005243092309683561, + 0.006055205129086971, + 0.004969994071871042, + 0.00547462422400713, + 0.01460554264485836, + 0.02034706436097622, + 0.027897844091057777, + 0.02145269326865673, + 0.02524220384657383, + 0.01807056926190853, + 0.015397859737277031, + 0.0164346881210804, + 0.016336878761649132, + 0.0203647892922163, + 0.02227158099412918, + 0.027944723144173622, + 0.012625457718968391, + 0.014779423363506794, + 0.014181427657604218, + 0.012238058261573315, + 0.015614977106451988, + 0.006239787209779024, + 0.005747986026108265, + 0.006014698185026646, + 0.00334780290722847, + 0.0039208997040987015, + 0.003953590057790279, + 0.003441530978307128, + 0.004177178721874952, + 0.012737398967146873, + 0.010946680791676044, + 0.016469726338982582, + 0.007514071185141802, + 0.006761555094271898, + 0.004901987500488758, + 0.005401913076639175, + 0.005220361519604921, + 0.01881852187216282, + 0.03280200809240341, + 0.016953006386756897, + 0.016750583425164223, + 0.017172962427139282, + 0.00862723495811224, + 0.015366737730801105, + 0.01296721026301384, + 0.011419197544455528, + 0.008557317778468132, + 0.010118436999619007, + 0.008608757518231869, + 0.0066588567569851875, + 0.0041458806954324245, + 0.008432075381278992, + 0.007264300249516964, + 0.004256104584783316, + 0.004842611961066723, + 0.004785806871950626, + 0.003991568461060524, + 0.004213037900626659, + 0.0015878146514296532, + 0.003532224101945758, + 0.003050177125260234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.08775977045297623 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821013167500496 + ], + "opposite_logprob_tokens": [ + -4.018209934234619 + ], + "logprob_reward": -0.01821013167500496, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01821013167500496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1310720145702362, + 0.18260358273983002, + 0.07426179200410843, + 0.03791869431734085, + 0.04148316755890846, + 0.09271611273288727, + 0.03637801483273506, + 0.11525742709636688, + 0.26354771852493286, + 0.024761470034718513 + ], + "credit_map": [ + 0.021661043167114258, + 0.015181804075837135, + 0.018200242891907692, + 0.0199873149394989, + 0.014332748018205166, + 0.010597381740808487, + 0.018890397623181343, + 0.012221071869134903, + 0.03992089629173279, + 0.037443872541189194, + 0.02715892158448696, + 0.016887567937374115, + 0.021710814908146858, + 0.009329454973340034, + 0.015999168157577515, + 0.014152868650853634, + 0.014641307294368744, + 0.0065633682534098625, + 0.0066325548104941845, + 0.0061170305125415325, + 0.010148806497454643, + 0.006342298351228237, + 0.007375250570476055, + 0.016441170126199722, + 0.005743889603763819, + 0.008341946639120579, + 0.005823825486004353, + 0.0051814294420182705, + 0.00395653210580349, + 0.0026335048023611307, + 0.003551628440618515, + 0.0026859338395297527, + 0.005301815923303366, + 0.007344587240368128, + 0.005377392750233412, + 0.004646429792046547, + 0.005732969846576452, + 0.004598848056048155, + 0.004183923359960318, + 0.004297194536775351, + 0.014559776522219181, + 0.016225818544626236, + 0.015295158140361309, + 0.010345445945858955, + 0.008681954815983772, + 0.007712730672210455, + 0.00974369514733553, + 0.010151518508791924, + 0.0056127323769032955, + 0.004969761706888676, + 0.0032115550711750984, + 0.005386112257838249, + 0.004752559587359428, + 0.0032458179630339146, + 0.0033380857203155756, + 0.005861389916390181, + 0.015682002529501915, + 0.020732853561639786, + 0.01264537125825882, + 0.016055596992373466, + 0.01603694073855877, + 0.007482919376343489, + 0.0144279720261693, + 0.012193764559924603, + 0.05548432096838951, + 0.06477832049131393, + 0.03350990265607834, + 0.028226889669895172, + 0.028854021802544594, + 0.016150549054145813, + 0.017269814386963844, + 0.019273873418569565, + 0.003997283987700939, + 0.005062334705144167, + 0.0048277731984853745, + 0.002344616688787937, + 0.0025688253808766603, + 0.0020521129481494427, + 0.002042762702330947, + 0.0018657599575817585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.01821013167500496 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203292317688465 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014203292317688465, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203292317688465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08583085238933563, + 0.06615374982357025, + 0.3612491488456726, + 0.026701875030994415, + 0.058583252131938934, + 0.03110962174832821, + 0.025379711762070656, + 0.3245229125022888, + 0.01509876362979412, + 0.005370180122554302 + ], + "credit_map": [ + 0.02397644892334938, + 0.010415532626211643, + 0.01948878914117813, + 0.010057773441076279, + 0.005286431405693293, + 0.006197675131261349, + 0.0043777176178991795, + 0.0060304817743599415, + 0.01479024812579155, + 0.014356969855725765, + 0.009621119126677513, + 0.005475349258631468, + 0.007417028304189444, + 0.0037498390302062035, + 0.005758447106927633, + 0.004984751343727112, + 0.08442972600460052, + 0.04401912912726402, + 0.08664366602897644, + 0.03397909179329872, + 0.028250930830836296, + 0.027736205607652664, + 0.02180519886314869, + 0.03438520058989525, + 0.0047593084163963795, + 0.004728805273771286, + 0.0031573346350342035, + 0.0031869402155280113, + 0.003322090022265911, + 0.0019112578593194485, + 0.0027453191578388214, + 0.0028908210806548595, + 0.008360193111002445, + 0.008409165777266026, + 0.011649004183709621, + 0.0049054245464503765, + 0.0057482109405100346, + 0.006890956312417984, + 0.0049862731248140335, + 0.007634023204445839, + 0.004624509252607822, + 0.005099611356854439, + 0.0047352127730846405, + 0.0035874242894351482, + 0.0025155546609312296, + 0.002948388224467635, + 0.003054471453651786, + 0.004544450435787439, + 0.003916600719094276, + 0.004413458984345198, + 0.005941150244325399, + 0.0035637947730720043, + 0.0025885230861604214, + 0.0013953199377283454, + 0.0017925886204466224, + 0.001768273301422596, + 0.04536670073866844, + 0.07612831890583038, + 0.03926333412528038, + 0.03780978545546532, + 0.04039255902171135, + 0.01940329745411873, + 0.0358009859919548, + 0.030357923358678818, + 0.0029381185304373503, + 0.0029216420371085405, + 0.0027670911513268948, + 0.0014064592542126775, + 0.00168910576030612, + 0.0010386472567915916, + 0.0010613765334710479, + 0.00127632194198668, + 0.0009149585966952145, + 0.0005925817531533539, + 0.0006153304129838943, + 0.0005852480535395443, + 0.000764854543376714, + 0.00043860869482159615, + 0.000765593140386045, + 0.0006930044619366527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.014203292317688465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944311618804932 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.005944311618804932, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944311618804932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11393653601408005, + 0.18294721841812134, + 0.10114120692014694, + 0.10676887631416321, + 0.09920667856931686, + 0.04943546652793884, + 0.06577855348587036, + 0.1428142786026001, + 0.11401300877332687, + 0.023958152160048485 + ], + "credit_map": [ + 0.024774610996246338, + 0.01105490978807211, + 0.01417815126478672, + 0.012586597353219986, + 0.012394244782626629, + 0.011512691155076027, + 0.01461376715451479, + 0.012821564450860023, + 0.030011165887117386, + 0.04582035169005394, + 0.037068046629428864, + 0.018543966114521027, + 0.012651396915316582, + 0.010363217443227768, + 0.014534308575093746, + 0.013954772613942623, + 0.02519424445927143, + 0.006878250744193792, + 0.01713421568274498, + 0.007992586120963097, + 0.009093553759157658, + 0.009104221127927303, + 0.006583309266716242, + 0.01916082575917244, + 0.01638471893966198, + 0.02123102732002735, + 0.015069427900016308, + 0.012626002542674541, + 0.009426315315067768, + 0.009967327117919922, + 0.00796809233725071, + 0.01409595925360918, + 0.013910233043134212, + 0.017928728833794594, + 0.014360249973833561, + 0.010777207091450691, + 0.01103876531124115, + 0.010375471785664558, + 0.008172514848411083, + 0.012643499299883842, + 0.007670908700674772, + 0.009139702655375004, + 0.0064244563691318035, + 0.0055860597640275955, + 0.005586135201156139, + 0.004571572877466679, + 0.005223326850682497, + 0.005233307369053364, + 0.008198248222470284, + 0.01429867371916771, + 0.008156237192451954, + 0.009198247455060482, + 0.008611932396888733, + 0.004946636036038399, + 0.007376397494226694, + 0.004992183297872543, + 0.019306596368551254, + 0.020881090313196182, + 0.015244273468852043, + 0.01623685471713543, + 0.019417444244027138, + 0.013500488363206387, + 0.02124333567917347, + 0.016984188929200172, + 0.017635084688663483, + 0.020157095044851303, + 0.016652340069413185, + 0.011905582621693611, + 0.01563914306461811, + 0.007810965180397034, + 0.012467455118894577, + 0.011745339259505272, + 0.003627280006185174, + 0.003825544845312834, + 0.0040558925829827785, + 0.0026745132636278868, + 0.002398218261078, + 0.0023282684851437807, + 0.002204683143645525, + 0.0028437520377337933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.005944311618804932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a heavy, polished stainless steel meat grinder clamped to a light-colored granite countertop visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a human hand using a cylindrical wooden plunger to push raw meat into the grinder's hopper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is coarse, deep-red ground meat accumulating in a deep, shiny stainless steel bowl below the grinder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background softly blurred showing wooden cabinets, a stainless steel refrigerator, a ceramic knife block, and a small potted herb plant?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069553904235363 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.011069553904235363, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069553904235363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12550215423107147, + 0.10819113999605179, + 0.16258497536182404, + 0.15946875512599945, + 0.05498599261045456, + 0.03624550625681877, + 0.05275462940335274, + 0.15876159071922302, + 0.12558183073997498, + 0.015923449769616127 + ], + "credit_map": [ + 0.019611461088061333, + 0.016164857894182205, + 0.017048006877303123, + 0.021587656810879707, + 0.014885487034916878, + 0.01364239864051342, + 0.011163443326950073, + 0.011398841626942158, + 0.01771196536719799, + 0.029308291152119637, + 0.016839072108268738, + 0.011663910001516342, + 0.009706043638288975, + 0.005615655332803726, + 0.008034566417336464, + 0.009311642497777939, + 0.0420447513461113, + 0.01393149234354496, + 0.03741447627544403, + 0.012240380048751831, + 0.012359497137367725, + 0.013425388373434544, + 0.009329507127404213, + 0.021839482709765434, + 0.021152323111891747, + 0.03638944402337074, + 0.02322974242269993, + 0.023454222828149796, + 0.018083486706018448, + 0.010326444171369076, + 0.015680259093642235, + 0.011152828112244606, + 0.006805982906371355, + 0.009579540230333805, + 0.00838907714933157, + 0.005810500588268042, + 0.00471658306196332, + 0.006844226270914078, + 0.00642376160249114, + 0.006416323594748974, + 0.006969992537051439, + 0.005503405351191759, + 0.004823275376111269, + 0.00307790981605649, + 0.0035248377826064825, + 0.003337075700983405, + 0.003980440087616444, + 0.005028569605201483, + 0.009060695767402649, + 0.006795825902372599, + 0.013754358515143394, + 0.005633073393255472, + 0.005458966363221407, + 0.0031123869121074677, + 0.005191216245293617, + 0.003748109797015786, + 0.02097299136221409, + 0.0332171656191349, + 0.01830301620066166, + 0.019274136051535606, + 0.02282085455954075, + 0.010238646529614925, + 0.019988829270005226, + 0.01394596416503191, + 0.02374914661049843, + 0.02693750336766243, + 0.028076058253645897, + 0.013195524923503399, + 0.011385162360966206, + 0.006652222014963627, + 0.006830610800534487, + 0.008755599148571491, + 0.0024286501575261354, + 0.001854295376688242, + 0.0014174957759678364, + 0.002470852807164192, + 0.001825062558054924, + 0.0016133838798850775, + 0.0024370651226490736, + 0.0018766435096040368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.011069553904235363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_0/index_tensor([6620, 6620])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_0/index_tensor([6620, 6620])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..123a578889d9ff0ea6306f9bf1ddce14afb9edcd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_0/index_tensor([6620, 6620])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme, eye-level, static macro shot captures a standard cardboard milk carton placed on a dark, polished granite kitchen countertop. The carton's waxy surface shows faint condensation, and a tiny, almost imperceptible hole is visible at the bottom. From this opening, a single, perfectly opaque droplet of white milk meticulously expands, its surface shimmering with the soft, directed stream of morning light from an unseen window. Driven by surface tension, the droplet breaks free in a tense moment, slowly and deliberately falling through the still air. The camera maintains sharp focus as the droplet impacts the smooth, cool surface of the counter, shattering into a delicate, crown-shaped splash in a microsecond before quickly collapsing and spreading into a small, irregular, milky puddle. The shallow depth of field blurs the background into a soft bokeh that obscures the shape of the kitchen, creating a quiet, contemplative atmosphere that isolates and elevates this minute, dramatic event.", + "seed": 1567, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_0/index_6620_seed_1567_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02182851326651871, + "vlm_logprob": -0.02182851326651871, + "vlm_yes_no_margin": 4.374999952316284, + "vlm_token_credit": -0.02182851326651871, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.057444170117378235 + ], + "opposite_logprob_tokens": [ + -2.9324440956115723 + ], + "logprob_reward": -0.057444170117378235, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.057444170117378235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14118817448616028, + 0.20934709906578064, + 0.10730774700641632, + 0.04715453460812569, + 0.15151330828666687, + 0.11759128421545029, + 0.07302944362163544, + 0.05708003416657448, + 0.06985063850879669, + 0.025937743484973907 + ], + "credit_map": [ + 0.014427736401557922, + 0.014271560125052929, + 0.014960628002882004, + 0.04699906334280968, + 0.007800975814461708, + 0.011296276934444904, + 0.017896456643939018, + 0.013535472564399242, + 0.015915118157863617, + 0.019738947972655296, + 0.025686562061309814, + 0.04814746230840683, + 0.023075448349118233, + 0.026265231892466545, + 0.028300995007157326, + 0.022217322140932083, + 0.009177555330097675, + 0.0117515679448843, + 0.015729155391454697, + 0.008688357658684254, + 0.011955862864851952, + 0.013973863795399666, + 0.010950559750199318, + 0.02508082240819931, + 0.007044544443488121, + 0.006038701627403498, + 0.007054040674120188, + 0.005941150244325399, + 0.0049734171479940414, + 0.005912089254707098, + 0.005707309581339359, + 0.004483281169086695, + 0.013860764913260937, + 0.013459274545311928, + 0.027695799246430397, + 0.013197413645684719, + 0.01724524237215519, + 0.016488559544086456, + 0.012729545123875141, + 0.036836713552474976, + 0.011854400858283043, + 0.011280889622867107, + 0.01989949867129326, + 0.017197055742144585, + 0.010124985128641129, + 0.008491506800055504, + 0.012252303771674633, + 0.02649064175784588, + 0.006584100890904665, + 0.007256707176566124, + 0.010333596728742123, + 0.008245309814810753, + 0.007385553326457739, + 0.007380093447864056, + 0.00959798227995634, + 0.01624610275030136, + 0.007246171124279499, + 0.007694230880588293, + 0.008022397756576538, + 0.004902671091258526, + 0.009195948019623756, + 0.0075043318793177605, + 0.0060057505033910275, + 0.0065085371024906635, + 0.011337240226566792, + 0.00883316807448864, + 0.010484042577445507, + 0.006118888035416603, + 0.007582367863506079, + 0.008635367266833782, + 0.0069874650798738, + 0.009872101247310638, + 0.0034853473771363497, + 0.003352066734805703, + 0.004235525149852037, + 0.002619929378852248, + 0.0036627757363021374, + 0.0034700275864452124, + 0.0019249088363721967, + 0.00318716443143785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.057444170117378235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018203457817435265 + ], + "opposite_logprob_tokens": [ + -4.018203258514404 + ], + "logprob_reward": -0.018203457817435265, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018203457817435265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19974474608898163, + 0.04650811105966568, + 0.1276375949382782, + 0.3351587653160095, + 0.09436279535293579, + 0.035929497331380844, + 0.0447438545525074, + 0.04802364110946655, + 0.03880392014980316, + 0.029087042436003685 + ], + "credit_map": [ + 0.011524958536028862, + 0.03086029179394245, + 0.019493959844112396, + 0.06696919351816177, + 0.014516490511596203, + 0.021473638713359833, + 0.0186721533536911, + 0.016234058886766434, + 0.0025639908853918314, + 0.005307814572006464, + 0.00514940544962883, + 0.00948201585561037, + 0.008673690259456635, + 0.0028573099989444017, + 0.007043283898383379, + 0.005430601537227631, + 0.018592076376080513, + 0.022132955491542816, + 0.017115827649831772, + 0.013672500848770142, + 0.01719851605594158, + 0.01487563643604517, + 0.01141128409653902, + 0.012638799846172333, + 0.025721071287989616, + 0.028218798339366913, + 0.051277294754981995, + 0.03365462273359299, + 0.06725937128067017, + 0.0241390373557806, + 0.043067075312137604, + 0.06182152032852173, + 0.007718215696513653, + 0.006564284209161997, + 0.01571298949420452, + 0.007691464386880398, + 0.019278310239315033, + 0.009361812844872475, + 0.008489618077874184, + 0.01954609900712967, + 0.003682406386360526, + 0.003919421695172787, + 0.006061718333512545, + 0.006037368904799223, + 0.003437486942857504, + 0.0029532420448958874, + 0.004011101089417934, + 0.005826752167195082, + 0.005362262018024921, + 0.005599331576377153, + 0.006434153765439987, + 0.0048600612208247185, + 0.004432907793670893, + 0.004579720087349415, + 0.005267093423753977, + 0.008208325132727623, + 0.007577461656183004, + 0.008437209762632847, + 0.0074575296603143215, + 0.0034928543027490377, + 0.008027528412640095, + 0.004454774782061577, + 0.004529474303126335, + 0.004046807065606117, + 0.006560564041137695, + 0.005175692494958639, + 0.005749711766839027, + 0.0034845659974962473, + 0.0039285933598876, + 0.003794108284637332, + 0.003173599950969219, + 0.006937086582183838, + 0.006301466375589371, + 0.0058281258679926395, + 0.004630441311746836, + 0.0023689165245741606, + 0.002227229531854391, + 0.0025421888567507267, + 0.0013445131480693817, + 0.0038441591896116734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.018203457817435265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361288502812386 + ], + "opposite_logprob_tokens": [ + -3.6513612270355225 + ], + "logprob_reward": -0.026361288502812386, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361288502812386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2215014547109604, + 0.21107769012451172, + 0.18743933737277985, + 0.08725884556770325, + 0.08787058293819427, + 0.059083275496959686, + 0.04121025279164314, + 0.03177826479077339, + 0.055181462317705154, + 0.017598800361156464 + ], + "credit_map": [ + 0.01176845096051693, + 0.041305143386125565, + 0.021999038755893707, + 0.0829455554485321, + 0.014421125873923302, + 0.013281730934977531, + 0.02169228158891201, + 0.014088128693401814, + 0.013939796946942806, + 0.01960212178528309, + 0.025312429293990135, + 0.05741643160581589, + 0.028819311410188675, + 0.026392705738544464, + 0.027514224871993065, + 0.012080667540431023, + 0.02565382607281208, + 0.027381008490920067, + 0.02613467536866665, + 0.0224918145686388, + 0.02603255584836006, + 0.026700101792812347, + 0.015432422049343586, + 0.017612947151064873, + 0.007584656588733196, + 0.009131228551268578, + 0.012091648764908314, + 0.006914753001183271, + 0.019676707684993744, + 0.010189785622060299, + 0.010462384670972824, + 0.011207680217921734, + 0.00926093477755785, + 0.013457071967422962, + 0.011989722028374672, + 0.009263470768928528, + 0.016291284933686256, + 0.006037762388586998, + 0.007425840012729168, + 0.014144494198262691, + 0.004931494127959013, + 0.0057740057818591595, + 0.008722937665879726, + 0.007885098457336426, + 0.004694701172411442, + 0.007054944988340139, + 0.008983917534351349, + 0.011036179028451443, + 0.006315604783594608, + 0.005676134489476681, + 0.006736632902175188, + 0.0038350848481059074, + 0.0041094026528298855, + 0.0038564465939998627, + 0.0052902670577168465, + 0.0053906817920506, + 0.0037877571303397417, + 0.0035190426278859377, + 0.00514515396207571, + 0.0026617562398314476, + 0.0037326097954064608, + 0.003520495491102338, + 0.004644630942493677, + 0.004766818601638079, + 0.008527174592018127, + 0.007602033205330372, + 0.008321259170770645, + 0.005903854500502348, + 0.006300065200775862, + 0.005965237971395254, + 0.005016934592276812, + 0.007544904015958309, + 0.0028699487447738647, + 0.002516617998480797, + 0.0029087781440466642, + 0.0017856258200481534, + 0.0017088069580495358, + 0.0015880261780694127, + 0.001117584528401494, + 0.003103413386270404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026361288502812386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002490278100594878 + ], + "opposite_logprob_tokens": [ + -6.002490043640137 + ], + "logprob_reward": -0.002490278100594878, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002490278100594878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4492674469947815, + 0.054467398673295975, + 0.10147061944007874, + 0.07508497685194016, + 0.040793634951114655, + 0.09601391851902008, + 0.052906766533851624, + 0.061532191932201385, + 0.03346940502524376, + 0.03499370813369751 + ], + "credit_map": [ + 0.024384740740060806, + 0.08164901286363602, + 0.040782541036605835, + 0.15149058401584625, + 0.040585778653621674, + 0.037258394062519073, + 0.045823682099580765, + 0.027292700484395027, + 0.0030415402725338936, + 0.006419513374567032, + 0.00660122511908412, + 0.00799141637980938, + 0.0075846086256206036, + 0.006618201266974211, + 0.006403375416994095, + 0.009807517752051353, + 0.011727885343134403, + 0.016390599310398102, + 0.012319788336753845, + 0.012558062560856342, + 0.013745962642133236, + 0.013510305434465408, + 0.008628513664007187, + 0.012589499354362488, + 0.006190177518874407, + 0.00887767132371664, + 0.013332834467291832, + 0.008330858312547207, + 0.012116924859583378, + 0.005151693243533373, + 0.008848896250128746, + 0.012235919013619423, + 0.003955418709665537, + 0.0034603162202984095, + 0.005863522179424763, + 0.005121207796037197, + 0.007117696106433868, + 0.0036541949957609177, + 0.006957581266760826, + 0.004663699306547642, + 0.009210806339979172, + 0.011303449980914593, + 0.01761670783162117, + 0.01403508149087429, + 0.008914394304156303, + 0.007149414625018835, + 0.010748474858701229, + 0.0170355886220932, + 0.005366244353353977, + 0.006493351887911558, + 0.008223684504628181, + 0.005918286275118589, + 0.0062937503680586815, + 0.004937160760164261, + 0.005604000296443701, + 0.010070287622511387, + 0.012923634611070156, + 0.01048852875828743, + 0.009145362302660942, + 0.00444367341697216, + 0.004323652479797602, + 0.00579063082113862, + 0.0065336404368281364, + 0.007883070036768913, + 0.003958877641707659, + 0.004023869056254625, + 0.00495366845279932, + 0.004543221555650234, + 0.0031591830775141716, + 0.0030681246425956488, + 0.004251517355442047, + 0.00551094114780426, + 0.004782878793776035, + 0.005963152274489403, + 0.006219570059329271, + 0.004394953139126301, + 0.0036351613234728575, + 0.002585494425147772, + 0.002923838561400771, + 0.004488658159971237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.002490278100594878 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643371794372797 + ], + "opposite_logprob_tokens": [ + -5.379643440246582 + ], + "logprob_reward": -0.004643371794372797, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643371794372797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2546997368335724, + 0.08897827565670013, + 0.06940806657075882, + 0.08465896546840668, + 0.1665104478597641, + 0.1500060111284256, + 0.07198385894298553, + 0.07968169450759888, + 0.017757846042513847, + 0.01631506159901619 + ], + "credit_map": [ + 0.01521261502057314, + 0.04411535710096359, + 0.022748375311493874, + 0.08050662279129028, + 0.023162439465522766, + 0.025965342298150063, + 0.026213685050606728, + 0.016775302588939667, + 0.006388538517057896, + 0.0074478210881352425, + 0.01010769885033369, + 0.021469712257385254, + 0.011021948419511318, + 0.013458020985126495, + 0.01178282406181097, + 0.007301718927919865, + 0.006030698772519827, + 0.009048684500157833, + 0.009108340367674828, + 0.008846983313560486, + 0.009109236299991608, + 0.012313506565988064, + 0.006878400221467018, + 0.008072219789028168, + 0.011479114182293415, + 0.005263231694698334, + 0.012947963550686836, + 0.008165485225617886, + 0.006802968680858612, + 0.012665623798966408, + 0.010412584990262985, + 0.0169220007956028, + 0.014609367586672306, + 0.011009447276592255, + 0.026337655261158943, + 0.01586814597249031, + 0.015836434438824654, + 0.01702798157930374, + 0.023582370951771736, + 0.04223904386162758, + 0.01385960541665554, + 0.017003843560814857, + 0.02520136348903179, + 0.02053372748196125, + 0.013701445423066616, + 0.013656486757099628, + 0.014013934880495071, + 0.03203560411930084, + 0.013101151213049889, + 0.01087103970348835, + 0.013594826683402061, + 0.007371979299932718, + 0.006276566069573164, + 0.006362269166857004, + 0.006390433292835951, + 0.008015592582523823, + 0.00767781026661396, + 0.009342556819319725, + 0.01268252357840538, + 0.007141521200537682, + 0.008197265677154064, + 0.011197799816727638, + 0.013544684275984764, + 0.009897527284920216, + 0.004195345100015402, + 0.002898941747844219, + 0.0029349352698773146, + 0.0010020708432421088, + 0.0010735709220170975, + 0.0012960234889760613, + 0.0010972382733598351, + 0.0032597207464277744, + 0.001545847044326365, + 0.002044833265244961, + 0.0025595766492187977, + 0.0017113768262788653, + 0.0020744511857628822, + 0.0021837232634425163, + 0.001160024432465434, + 0.003035229165107012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.004643371794372797 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme, eye-level, static macro shot captures a standard cardboard milk carton placed on a dark, polished granite kitchen countertop. The carton's waxy surface shows faint condensation, and a tiny, almost imperceptible hole is visible at the bottom. From this opening, a single, perfectly opaque droplet of white milk meticulously expands, its surface shimmering with the soft, directed stream of morning light from an unseen window. Driven by surface tension, the droplet breaks free in a tense moment, slowly and deliberately falling through the still air. The camera maintains sharp focus as the droplet impacts the smooth, cool surface of the counter, shattering into a delicate, crown-shaped splash in a microsecond before quickly collapsing and spreading into a small, irregular, milky puddle. The shallow depth of field blurs the background into a soft bokeh that obscures the shape of the kitchen, creating a quiet, contemplative atmosphere that isolates and elevates this minute, dramatic event.", + "seed": 1568, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_0/index_6620_seed_1568_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.027302035316824912, + "vlm_logprob": -0.027302035316824912, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.027302035316824912, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07180414348840714 + ], + "opposite_logprob_tokens": [ + -2.8218040466308594 + ], + "logprob_reward": -0.07180414348840714, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07180414348840714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04827002435922623, + 0.05924483388662338, + 0.02681562677025795, + 0.04974259063601494, + 0.2878282070159912, + 0.06936787813901901, + 0.06077469512820244, + 0.19804158806800842, + 0.1290373057126999, + 0.07087729126214981 + ], + "credit_map": [ + 0.007705779280513525, + 0.004697229713201523, + 0.012724004685878754, + 0.007408826611936092, + 0.004593429155647755, + 0.004355678800493479, + 0.0034192530438303947, + 0.003365815617144108, + 0.009520416148006916, + 0.00852061528712511, + 0.008241473697125912, + 0.007798412814736366, + 0.009984796866774559, + 0.0033036291133612394, + 0.006583529524505138, + 0.005291954148560762, + 0.0034953707363456488, + 0.00514969602227211, + 0.00282529485411942, + 0.003538387129083276, + 0.003914147149771452, + 0.0014926090370863676, + 0.003373314393684268, + 0.003026803722605109, + 0.007976839318871498, + 0.010226179845631123, + 0.004790551029145718, + 0.008137371391057968, + 0.003146334085613489, + 0.0037722967099398375, + 0.005312125664204359, + 0.0063808877021074295, + 0.0566011518239975, + 0.07519491016864777, + 0.029538854956626892, + 0.020621325820684433, + 0.013488482683897018, + 0.021112842485308647, + 0.018187325447797775, + 0.05308329686522484, + 0.00888406578451395, + 0.012035749852657318, + 0.019818326458334923, + 0.007297429721802473, + 0.003340683411806822, + 0.00397739140316844, + 0.00481250649318099, + 0.009201718494296074, + 0.007906841114163399, + 0.010960598476231098, + 0.01206613052636385, + 0.006822521798312664, + 0.003105433424934745, + 0.004436842631548643, + 0.005831450689584017, + 0.009644865058362484, + 0.02613994851708412, + 0.041368186473846436, + 0.06408675760030746, + 0.022915614768862724, + 0.007743947207927704, + 0.011007741093635559, + 0.009496456943452358, + 0.015282919630408287, + 0.018423138186335564, + 0.02652338705956936, + 0.023134514689445496, + 0.012743075378239155, + 0.007165627554059029, + 0.01086273230612278, + 0.008452489040791988, + 0.021732326596975327, + 0.008695551194250584, + 0.012262372300028801, + 0.012959975749254227, + 0.010512121953070164, + 0.0034851885866373777, + 0.004433073569089174, + 0.012240498326718807, + 0.00628850469365716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.07180414348840714 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011099499650299549 + ], + "opposite_logprob_tokens": [ + -4.511099338531494 + ], + "logprob_reward": -0.011099499650299549, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011099499650299549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0923367291688919, + 0.06255358457565308, + 0.07370253652334213, + 0.26151883602142334, + 0.25948819518089294, + 0.03002243861556053, + 0.02841755375266075, + 0.034878697246313095, + 0.13349385559558868, + 0.02358754351735115 + ], + "credit_map": [ + 0.013416989706456661, + 0.010747765190899372, + 0.012386939488351345, + 0.010514850728213787, + 0.012355189770460129, + 0.013735041953623295, + 0.010800116695463657, + 0.008379836566746235, + 0.009946484118700027, + 0.009327792562544346, + 0.009609359316527843, + 0.006715319585055113, + 0.009239339269697666, + 0.004769972991198301, + 0.006258804351091385, + 0.006686513312160969, + 0.011136536486446857, + 0.012878157198429108, + 0.010689940303564072, + 0.008671515621244907, + 0.008516475558280945, + 0.005550338886678219, + 0.006631594616919756, + 0.009627974592149258, + 0.04059552401304245, + 0.060241516679525375, + 0.02401696890592575, + 0.040651045739650726, + 0.02316250465810299, + 0.018144454807043076, + 0.025292396545410156, + 0.02941441349685192, + 0.04960131645202637, + 0.06925056129693985, + 0.029584161937236786, + 0.021574873477220535, + 0.013254058547317982, + 0.01729419082403183, + 0.01473471149802208, + 0.04419431835412979, + 0.003466620109975338, + 0.004137504380196333, + 0.0038777042645961046, + 0.0033017240930348635, + 0.002300497144460678, + 0.0028463066555559635, + 0.003240286372601986, + 0.006851795595139265, + 0.004125454928725958, + 0.005209822673350573, + 0.00527039822191, + 0.002804639283567667, + 0.0016815116396173835, + 0.0030050084460526705, + 0.0022363706957548857, + 0.004084345418959856, + 0.003776548895984888, + 0.0048350621946156025, + 0.007685329299420118, + 0.004164691083133221, + 0.0016947014955803752, + 0.002666447777301073, + 0.005304665304720402, + 0.004751251079142094, + 0.017709970474243164, + 0.022217851132154465, + 0.02445441484451294, + 0.013709740713238716, + 0.0082017807289958, + 0.013868328183889389, + 0.010477975010871887, + 0.022853795439004898, + 0.0025415339041501284, + 0.002921880455687642, + 0.003883452620357275, + 0.0033115982078015804, + 0.0013374427799135447, + 0.0018661186331883073, + 0.004765734542161226, + 0.002959783188998699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.011099499650299549 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810933396220207 + ], + "opposite_logprob_tokens": [ + -3.529810905456543 + ], + "logprob_reward": -0.029810933396220207, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810933396220207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11836157739162445, + 0.10471272468566895, + 0.09763521701097488, + 0.18376345932483673, + 0.17060214281082153, + 0.06469583511352539, + 0.04961877316236496, + 0.05804939568042755, + 0.11618675291538239, + 0.036374084651470184 + ], + "credit_map": [ + 0.017004745081067085, + 0.012708522379398346, + 0.017369303852319717, + 0.01387422252446413, + 0.01577778160572052, + 0.01784503646194935, + 0.014605995267629623, + 0.009175989776849747, + 0.017916439101099968, + 0.022883016616106033, + 0.014897349290549755, + 0.012429826892912388, + 0.013864659704267979, + 0.006115655414760113, + 0.008472172543406487, + 0.008133613504469395, + 0.012245032005012035, + 0.017014024779200554, + 0.011138537898659706, + 0.013683381490409374, + 0.009021840989589691, + 0.008229334838688374, + 0.01521498616784811, + 0.011088096536695957, + 0.033326052129268646, + 0.04193120449781418, + 0.021162282675504684, + 0.020511310547590256, + 0.015351526439189911, + 0.011693448759615421, + 0.020495528355240822, + 0.019292127341032028, + 0.03418136015534401, + 0.04585328325629234, + 0.018670611083507538, + 0.012347391806542873, + 0.00882435031235218, + 0.01213166955858469, + 0.00999151449650526, + 0.028601985424757004, + 0.008008765056729317, + 0.009480404667556286, + 0.011098350398242474, + 0.006799423135817051, + 0.005275765899568796, + 0.0065924013033509254, + 0.008761601522564888, + 0.008679121732711792, + 0.005456554237753153, + 0.007290721870958805, + 0.008774837478995323, + 0.005313689820468426, + 0.0021542897447943687, + 0.004021424800157547, + 0.00551129225641489, + 0.011095969006419182, + 0.012275643646717072, + 0.015006925910711288, + 0.006632860749959946, + 0.004951280541718006, + 0.004661685787141323, + 0.004475281573832035, + 0.005982090253382921, + 0.004063635598868132, + 0.017879758030176163, + 0.019742533564567566, + 0.01924692653119564, + 0.014191881753504276, + 0.007458196487277746, + 0.012633762322366238, + 0.01034634280949831, + 0.014687364920973778, + 0.005001893267035484, + 0.004863785579800606, + 0.00609038770198822, + 0.005139932036399841, + 0.001892486703582108, + 0.0033150268718600273, + 0.005138405133038759, + 0.004932173527777195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.029810933396220207 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195776604115963 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.003195776604115963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195776604115963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07308803498744965, + 0.092253178358078, + 0.03355005756020546, + 0.07731880247592926, + 0.3121091425418854, + 0.033091992139816284, + 0.026248713955283165, + 0.07895911484956741, + 0.10899293422698975, + 0.1643880307674408 + ], + "credit_map": [ + 0.009962128475308418, + 0.009662269614636898, + 0.010509125888347626, + 0.00694911228492856, + 0.009565399959683418, + 0.01183203887194395, + 0.007862253114581108, + 0.006745703052729368, + 0.012820003554224968, + 0.012625455856323242, + 0.012307118624448776, + 0.011484894901514053, + 0.012768860906362534, + 0.006743844598531723, + 0.0152504937723279, + 0.008252492174506187, + 0.0049180928617715836, + 0.005833376664668322, + 0.0029720659367740154, + 0.004008900839835405, + 0.003359900787472725, + 0.002877300139516592, + 0.004056701902300119, + 0.005523716099560261, + 0.01059700082987547, + 0.015275498852133751, + 0.007446255534887314, + 0.012794204987585545, + 0.007887323386967182, + 0.005458472296595573, + 0.009664036333560944, + 0.00819600559771061, + 0.06299486756324768, + 0.08382264524698257, + 0.03391565382480621, + 0.022926993668079376, + 0.015734070912003517, + 0.022142162546515465, + 0.016615182161331177, + 0.05395754054188728, + 0.004902053624391556, + 0.0060001155361533165, + 0.007556830998510122, + 0.004196415655314922, + 0.0019442352931946516, + 0.0025566210970282555, + 0.0033398899249732494, + 0.0025958262849599123, + 0.00356596359051764, + 0.004098644480109215, + 0.005166750866919756, + 0.00267412094399333, + 0.0015065100742504, + 0.0033295319881290197, + 0.002334584016352892, + 0.003572608344256878, + 0.012284203432500362, + 0.01679876260459423, + 0.017381593585014343, + 0.007849547080695629, + 0.003851588349789381, + 0.004757790360599756, + 0.006692706141620874, + 0.00934291910380125, + 0.015241365879774094, + 0.017062455415725708, + 0.019426874816417694, + 0.010346607305109501, + 0.007266611326485872, + 0.014152674935758114, + 0.008266065269708633, + 0.01723027043044567, + 0.01995144598186016, + 0.02812308631837368, + 0.028820473700761795, + 0.02550947479903698, + 0.008190884254872799, + 0.01051499042659998, + 0.02892572246491909, + 0.01435194630175829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.003195776604115963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059982344508171 + ], + "opposite_logprob_tokens": [ + -3.895599842071533 + ], + "logprob_reward": -0.02059982344508171, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059982344508171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0474027618765831, + 0.05093195289373398, + 0.026222649961709976, + 0.019403623417019844, + 0.44301527738571167, + 0.02453341893851757, + 0.021597247570753098, + 0.03539281710982323, + 0.12870201468467712, + 0.20279817283153534 + ], + "credit_map": [ + 0.009944302029907703, + 0.003903846023604274, + 0.014418759383261204, + 0.0072728111408650875, + 0.0027754998300224543, + 0.0026184057351201773, + 0.003971605561673641, + 0.0024975340347737074, + 0.007466307375580072, + 0.008275684900581837, + 0.009560951963067055, + 0.005947405938059092, + 0.00693862047046423, + 0.0030714271124452353, + 0.005733128637075424, + 0.003938428126275539, + 0.003560242708772421, + 0.005488136783242226, + 0.0023425286635756493, + 0.004896748811006546, + 0.003056940156966448, + 0.0021991601679474115, + 0.002131268149241805, + 0.0025476282462477684, + 0.0026816881727427244, + 0.003136258339509368, + 0.002185054123401642, + 0.003463351633399725, + 0.0015375680522993207, + 0.0015349832829087973, + 0.0029941690154373646, + 0.0018705520778894424, + 0.08440113812685013, + 0.11475086957216263, + 0.04740293323993683, + 0.03345491737127304, + 0.022287629544734955, + 0.032250821590423584, + 0.02565404586493969, + 0.08281298726797104, + 0.0027339530643075705, + 0.004058055579662323, + 0.007252111099660397, + 0.002990068169310689, + 0.001244292943738401, + 0.0016114915488287807, + 0.002229770878329873, + 0.0024136784486472607, + 0.00312259909696877, + 0.0041889469139277935, + 0.005736888386309147, + 0.0023308228701353073, + 0.0010022134520113468, + 0.0013102830853313208, + 0.0015305312117561698, + 0.00237496430054307, + 0.004490795079618692, + 0.006895069498568773, + 0.009273500181734562, + 0.004096829332411289, + 0.0015369838802143931, + 0.0024583213962614536, + 0.002990333829075098, + 0.0036509884521365166, + 0.019145885482430458, + 0.024942435324192047, + 0.022679178044199944, + 0.012031751684844494, + 0.007779621984809637, + 0.012667878530919552, + 0.00823949184268713, + 0.021215803921222687, + 0.023914163932204247, + 0.03343650698661804, + 0.036734141409397125, + 0.030731504783034325, + 0.0100591154769063, + 0.01278624590486288, + 0.0364910252392292, + 0.01864549145102501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.02059982344508171 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme, eye-level, static macro shot captures a standard cardboard milk carton placed on a dark, polished granite kitchen countertop. The carton's waxy surface shows faint condensation, and a tiny, almost imperceptible hole is visible at the bottom. From this opening, a single, perfectly opaque droplet of white milk meticulously expands, its surface shimmering with the soft, directed stream of morning light from an unseen window. Driven by surface tension, the droplet breaks free in a tense moment, slowly and deliberately falling through the still air. The camera maintains sharp focus as the droplet impacts the smooth, cool surface of the counter, shattering into a delicate, crown-shaped splash in a microsecond before quickly collapsing and spreading into a small, irregular, milky puddle. The shallow depth of field blurs the background into a soft bokeh that obscures the shape of the kitchen, creating a quiet, contemplative atmosphere that isolates and elevates this minute, dramatic event.", + "seed": 1569, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_0/index_6620_seed_1569_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.028685356304049493, + "vlm_logprob": -0.028685356304049493, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.028685356304049493, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07056985050439835 + ], + "opposite_logprob_tokens": [ + -2.8205697536468506 + ], + "logprob_reward": -0.07056985050439835, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07056985050439835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15245816111564636, + 0.03798766806721687, + 0.29498419165611267, + 0.1287766844034195, + 0.052359580993652344, + 0.05970877781510353, + 0.05593337491154671, + 0.11820605397224426, + 0.06491540372371674, + 0.034670133143663406 + ], + "credit_map": [ + 0.007614646572619677, + 0.0110430046916008, + 0.017637675628066063, + 0.019493427127599716, + 0.02329523116350174, + 0.025914154946804047, + 0.02833866886794567, + 0.019121354445815086, + 0.0020457925274968147, + 0.0030007055029273033, + 0.004776989109814167, + 0.0044512371532619, + 0.005460198037326336, + 0.005574439186602831, + 0.00715084420517087, + 0.0055274623446166515, + 0.022906653583049774, + 0.030226843431591988, + 0.03312228247523308, + 0.0420340932905674, + 0.05925492197275162, + 0.03563985228538513, + 0.042685311287641525, + 0.029114242643117905, + 0.009169030003249645, + 0.018728984519839287, + 0.011830355040729046, + 0.02543305605649948, + 0.011366372928023338, + 0.01473317015916109, + 0.02004002034664154, + 0.01747569441795349, + 0.0036164317280054092, + 0.0051863850094377995, + 0.00749474810436368, + 0.007407033815979958, + 0.005546271335333586, + 0.007259022444486618, + 0.006134736817330122, + 0.009714952670037746, + 0.006350765470415354, + 0.008381901308894157, + 0.007789726834744215, + 0.005197504535317421, + 0.008273574523627758, + 0.00894614402204752, + 0.004626178648322821, + 0.010142982937395573, + 0.0059511964209377766, + 0.008471122942864895, + 0.006257831584662199, + 0.007133233826607466, + 0.00958571769297123, + 0.00755939120426774, + 0.003969735000282526, + 0.007005148101598024, + 0.016134174540638924, + 0.012890645302832127, + 0.01356614101678133, + 0.01557004265487194, + 0.012128693051636219, + 0.015124750323593616, + 0.010609341785311699, + 0.022182265296578407, + 0.006295052822679281, + 0.006779975723475218, + 0.007704408839344978, + 0.009471266530454159, + 0.007082016207277775, + 0.00841866061091423, + 0.008532663807272911, + 0.010631353594362736, + 0.0022047546226531267, + 0.003518260782584548, + 0.004937979858368635, + 0.005947962403297424, + 0.003986394964158535, + 0.00473640626296401, + 0.00306770415045321, + 0.006270670332014561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.07056985050439835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638242021203041 + ], + "opposite_logprob_tokens": [ + -3.6513824462890625 + ], + "logprob_reward": -0.02638242021203041, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638242021203041, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19027574360370636, + 0.07156646996736526, + 0.28646180033683777, + 0.16086268424987793, + 0.04435352608561516, + 0.06807819753885269, + 0.029505519196391106, + 0.05695384740829468, + 0.05740419775247574, + 0.03453804925084114 + ], + "credit_map": [ + 0.012424927204847336, + 0.010664249770343304, + 0.01765846088528633, + 0.015109210275113583, + 0.02673027105629444, + 0.04879692941904068, + 0.04003556817770004, + 0.018856123089790344, + 0.0046303472481667995, + 0.00628929166123271, + 0.006405692547559738, + 0.006918405648320913, + 0.010241594165563583, + 0.013968980871140957, + 0.012385733425617218, + 0.010726421140134335, + 0.02038278989493847, + 0.028926849365234375, + 0.03335985168814659, + 0.03899899125099182, + 0.06305016577243805, + 0.03386126458644867, + 0.038241878151893616, + 0.029639998450875282, + 0.010333924554288387, + 0.016138847917318344, + 0.01597137749195099, + 0.031298212707042694, + 0.01058389525860548, + 0.019190777093172073, + 0.02995123900473118, + 0.027394406497478485, + 0.003655837383121252, + 0.0047214035876095295, + 0.005706955213099718, + 0.0061622001230716705, + 0.00532405823469162, + 0.005629777442663908, + 0.007355355657637119, + 0.005797938909381628, + 0.005325068719685078, + 0.010385903529822826, + 0.00844665803015232, + 0.0052625807002186775, + 0.009838842786848545, + 0.010187425650656223, + 0.005794509779661894, + 0.012837204150855541, + 0.0024027100298553705, + 0.0031409626826643944, + 0.004506842233240604, + 0.004126283340156078, + 0.0033071842044591904, + 0.003886423073709011, + 0.0034828423522412777, + 0.004652270581573248, + 0.006678575184196234, + 0.0061376020312309265, + 0.007650069426745176, + 0.010309896431863308, + 0.004005391616374254, + 0.007078931666910648, + 0.005371126811951399, + 0.009722252376377583, + 0.00589883653447032, + 0.0064312368631362915, + 0.005806670058518648, + 0.007606188300997019, + 0.006704261060804129, + 0.007535114884376526, + 0.00767684169113636, + 0.009745050221681595, + 0.0023564582224935293, + 0.004049558658152819, + 0.0038239466957747936, + 0.008056421764194965, + 0.0036296285688877106, + 0.004352721851319075, + 0.003373947227373719, + 0.004895368125289679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.02638242021203041 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029819844290614128 + ], + "opposite_logprob_tokens": [ + -3.5298197269439697 + ], + "logprob_reward": -0.029819844290614128, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029819844290614128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11629126220941544, + 0.1325937658548355, + 0.1506602019071579, + 0.2263914942741394, + 0.08500093966722488, + 0.07707912474870682, + 0.04468338564038277, + 0.10047388821840286, + 0.03752020746469498, + 0.02930581010878086 + ], + "credit_map": [ + 0.00681195268407464, + 0.011337735690176487, + 0.011362805962562561, + 0.013804749585688114, + 0.02192070335149765, + 0.020617375150322914, + 0.019180208444595337, + 0.011255736462771893, + 0.0055389404296875, + 0.010139379650354385, + 0.007923407480120659, + 0.013734063133597374, + 0.023034924641251564, + 0.02994729019701481, + 0.03333594277501106, + 0.008939811028540134, + 0.01109193917363882, + 0.014755831100046635, + 0.025024423375725746, + 0.01813182234764099, + 0.027836507186293602, + 0.01863463781774044, + 0.020883459597826004, + 0.014301585033535957, + 0.01831674762070179, + 0.03639313578605652, + 0.02092571184039116, + 0.04400717094540596, + 0.022369293496012688, + 0.026267847046256065, + 0.028376372531056404, + 0.029735201969742775, + 0.011088697239756584, + 0.012821977958083153, + 0.008523257449269295, + 0.011095470748841763, + 0.009570162743330002, + 0.012099905870854855, + 0.006394398398697376, + 0.013407072052359581, + 0.006247008219361305, + 0.012270137667655945, + 0.009002851322293282, + 0.005734618287533522, + 0.010968788526952267, + 0.011646490544080734, + 0.006585175171494484, + 0.014624056406319141, + 0.002690037712454796, + 0.006323665846139193, + 0.004810459446161985, + 0.006593701429665089, + 0.005561888217926025, + 0.006573636084794998, + 0.004552672617137432, + 0.0075773256830871105, + 0.010511939413845539, + 0.007946005091071129, + 0.010940657928586006, + 0.014735779725015163, + 0.008458287455141544, + 0.014158247038722038, + 0.009018812328577042, + 0.024704160168766975, + 0.002919689519330859, + 0.0032539430540055037, + 0.005228912923485041, + 0.00471850112080574, + 0.0027572661638259888, + 0.005328548606485128, + 0.003365296171978116, + 0.009948046877980232, + 0.002399967284873128, + 0.0019434051355347037, + 0.004700633697211742, + 0.004404010716825724, + 0.0021321431268006563, + 0.004472212865948677, + 0.002018433064222336, + 0.0072350045666098595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.029819844290614128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099300131201744 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099300131201744, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099300131201744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17729295790195465, + 0.05070468783378601, + 0.22386842966079712, + 0.18764667212963104, + 0.052929069846868515, + 0.04696049913764, + 0.06853422522544861, + 0.1284523904323578, + 0.03620876371860504, + 0.02740231715142727 + ], + "credit_map": [ + 0.008020726032555103, + 0.008112114854156971, + 0.014482343569397926, + 0.016029341146349907, + 0.025147274136543274, + 0.051445964723825455, + 0.03702631592750549, + 0.01702887937426567, + 0.0016949583077803254, + 0.002931106137111783, + 0.004686874803155661, + 0.005361654330044985, + 0.0056061833165585995, + 0.013899967074394226, + 0.012676393613219261, + 0.0038475501351058483, + 0.016781318932771683, + 0.02036166377365589, + 0.02690465748310089, + 0.03066927008330822, + 0.04626898467540741, + 0.025071971118450165, + 0.03307490795850754, + 0.024735648185014725, + 0.0124288285151124, + 0.02504671737551689, + 0.015032744966447353, + 0.04073620215058327, + 0.018802031874656677, + 0.018636228516697884, + 0.03311865031719208, + 0.023845259100198746, + 0.005694193299859762, + 0.0057143778540194035, + 0.005577593110501766, + 0.0060120634734630585, + 0.005987188778817654, + 0.007539676036685705, + 0.006599563639611006, + 0.00980441365391016, + 0.00465113390237093, + 0.007236256264150143, + 0.005378657951951027, + 0.003275746712461114, + 0.006992589216679335, + 0.007066231220960617, + 0.004338391125202179, + 0.008021493442356586, + 0.0056066750548779964, + 0.010127733461558819, + 0.006250231526792049, + 0.010201612487435341, + 0.011055530048906803, + 0.008677762001752853, + 0.007304003927856684, + 0.009310678578913212, + 0.01662381924688816, + 0.01333546917885542, + 0.013905505649745464, + 0.017017614096403122, + 0.012649504467844963, + 0.017014706507325172, + 0.012279041111469269, + 0.025626733899116516, + 0.0037314249202609062, + 0.003730644239112735, + 0.005105627700686455, + 0.0037087788805365562, + 0.0032655035611242056, + 0.005261101759970188, + 0.0034482665359973907, + 0.007957416586577892, + 0.0016841123579069972, + 0.0024763401597738266, + 0.003557190066203475, + 0.00534606259316206, + 0.00316600245423615, + 0.0036986535415053368, + 0.002499047201126814, + 0.004974908195436001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.004099300131201744 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255536638200283 + ], + "opposite_logprob_tokens": [ + -4.3875555992126465 + ], + "logprob_reward": -0.01255536638200283, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255536638200283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13947941362857819, + 0.034338437020778656, + 0.2695714831352234, + 0.10356872528791428, + 0.04043816030025482, + 0.039053741842508316, + 0.09283393621444702, + 0.179657980799675, + 0.055986881256103516, + 0.04507125914096832 + ], + "credit_map": [ + 0.005672964733093977, + 0.009429858066141605, + 0.014207825995981693, + 0.015335739590227604, + 0.02286829613149166, + 0.03043193742632866, + 0.026868047192692757, + 0.014664742164313793, + 0.0018428124021738768, + 0.00231171608902514, + 0.00443950854241848, + 0.003826945321634412, + 0.0033679346088320017, + 0.008244512602686882, + 0.007483979687094688, + 0.0028210298623889685, + 0.020354343578219414, + 0.027304086834192276, + 0.03056134469807148, + 0.03827659785747528, + 0.05273053050041199, + 0.03072994202375412, + 0.04001950845122337, + 0.029595104977488518, + 0.010046420618891716, + 0.013428623788058758, + 0.01010467205196619, + 0.017346490174531937, + 0.011245548725128174, + 0.011454080231487751, + 0.01512881275266409, + 0.014814077876508236, + 0.004150389693677425, + 0.004359542392194271, + 0.00511319050565362, + 0.004723331891000271, + 0.004389693029224873, + 0.0043778251856565475, + 0.007247262168675661, + 0.006076926365494728, + 0.0030002540443092585, + 0.004520853515714407, + 0.005296459887176752, + 0.004191866610199213, + 0.004949385300278664, + 0.005915425252169371, + 0.003461686661466956, + 0.007717813830822706, + 0.008513446897268295, + 0.01594158262014389, + 0.009324030950665474, + 0.012580337934195995, + 0.016543403267860413, + 0.011984612792730331, + 0.0077636572532355785, + 0.010182863101363182, + 0.023999135941267014, + 0.020020093768835068, + 0.020541062578558922, + 0.023113302886486053, + 0.019263241440057755, + 0.022837473079562187, + 0.017348570749163628, + 0.03253510966897011, + 0.00583341158926487, + 0.006447444204241037, + 0.006399518344551325, + 0.006546239368617535, + 0.006871117744594812, + 0.007317953743040562, + 0.007555339485406876, + 0.009015854448080063, + 0.0037143384106457233, + 0.004613457713276148, + 0.005555843468755484, + 0.007079549133777618, + 0.006164829712361097, + 0.0062661925330758095, + 0.004328332841396332, + 0.007348715327680111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.01255536638200283 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme, eye-level, static macro shot captures a standard cardboard milk carton placed on a dark, polished granite kitchen countertop. The carton's waxy surface shows faint condensation, and a tiny, almost imperceptible hole is visible at the bottom. From this opening, a single, perfectly opaque droplet of white milk meticulously expands, its surface shimmering with the soft, directed stream of morning light from an unseen window. Driven by surface tension, the droplet breaks free in a tense moment, slowly and deliberately falling through the still air. The camera maintains sharp focus as the droplet impacts the smooth, cool surface of the counter, shattering into a delicate, crown-shaped splash in a microsecond before quickly collapsing and spreading into a small, irregular, milky puddle. The shallow depth of field blurs the background into a soft bokeh that obscures the shape of the kitchen, creating a quiet, contemplative atmosphere that isolates and elevates this minute, dramatic event.", + "seed": 1570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_0/index_6620_seed_1570_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022546805813908576, + "vlm_logprob": -0.022546805813908576, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.022546805813908576, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032400939613580704 + ], + "opposite_logprob_tokens": [ + -4.032401084899902 + ], + "logprob_reward": -0.032400939613580704, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.032400939613580704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12381550669670105, + 0.08400244265794754, + 0.050970591604709625, + 0.0685986578464508, + 0.24276138842105865, + 0.1382281333208084, + 0.11181905120611191, + 0.10675905644893646, + 0.04955853521823883, + 0.023486550897359848 + ], + "credit_map": [ + 0.012730784714221954, + 0.01580744795501232, + 0.013434347696602345, + 0.019263435155153275, + 0.012519186362624168, + 0.015733571723103523, + 0.018087998032569885, + 0.016238750889897346, + 0.01073264330625534, + 0.007390663493424654, + 0.01939377561211586, + 0.010890903882682323, + 0.006884098518639803, + 0.008936310186982155, + 0.012505738064646721, + 0.007268316578119993, + 0.005826226435601711, + 0.00663762167096138, + 0.00612896541133523, + 0.0094534233212471, + 0.0060234349220991135, + 0.006778184324502945, + 0.006285564042627811, + 0.0038371742703020573, + 0.009792868047952652, + 0.008754223585128784, + 0.013709772378206253, + 0.009955755434930325, + 0.0057686385698616505, + 0.007052947301417589, + 0.006481329444795847, + 0.007083127740770578, + 0.026375431567430496, + 0.03087354451417923, + 0.052297480404376984, + 0.06315520405769348, + 0.01612202264368534, + 0.01263305265456438, + 0.018240714445710182, + 0.023063959553837776, + 0.017012838274240494, + 0.010745341889560223, + 0.02623695507645607, + 0.025878936052322388, + 0.011584142223000526, + 0.009440231136977673, + 0.021852383390069008, + 0.015477325767278671, + 0.019317004829645157, + 0.01367233507335186, + 0.014670834876596928, + 0.018904833123087883, + 0.012866612523794174, + 0.01147700846195221, + 0.01241288147866726, + 0.008497544564306736, + 0.018798191100358963, + 0.012488189153373241, + 0.020484820008277893, + 0.009939445182681084, + 0.01466530840843916, + 0.008657670579850674, + 0.011821171268820763, + 0.009904266335070133, + 0.008976948447525501, + 0.006201779004186392, + 0.010233026929199696, + 0.004814120940864086, + 0.004287554882466793, + 0.005525566171854734, + 0.004415076691657305, + 0.005104463547468185, + 0.0032130375038832426, + 0.0030397912487387657, + 0.002805954311043024, + 0.003444625297561288, + 0.002599874511361122, + 0.001981031149625778, + 0.003411302575841546, + 0.0029909361619502306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.032400939613580704 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005978795234113932 + ], + "opposite_logprob_tokens": [ + -5.130978584289551 + ], + "logprob_reward": -0.005978795234113932, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005978795234113932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10687471181154251, + 0.051336273550987244, + 0.04631972312927246, + 0.23374925553798676, + 0.06589704751968384, + 0.11081743985414505, + 0.1616286188364029, + 0.0853339210152626, + 0.10493779927492142, + 0.033105187118053436 + ], + "credit_map": [ + 0.01153344102203846, + 0.009324366226792336, + 0.026463663205504417, + 0.02100665308535099, + 0.007211992517113686, + 0.009242141619324684, + 0.011978516355156898, + 0.010113933123648167, + 0.007249058689922094, + 0.007155146449804306, + 0.007883408106863499, + 0.006650181487202644, + 0.006251196376979351, + 0.0064510139636695385, + 0.006021793000400066, + 0.0036744752433151007, + 0.00571045046672225, + 0.00466845091432333, + 0.008268939331173897, + 0.006954404059797525, + 0.003431744175031781, + 0.0038385556545108557, + 0.009767116978764534, + 0.003680066205561161, + 0.030425993725657463, + 0.03726387023925781, + 0.029941264539957047, + 0.030012881383299828, + 0.02591080032289028, + 0.014319811016321182, + 0.0493573397397995, + 0.016517290845513344, + 0.01079822052270174, + 0.007489215116947889, + 0.01264373492449522, + 0.01152003463357687, + 0.0055549489334225655, + 0.004663145635277033, + 0.008720134384930134, + 0.004507618490606546, + 0.013692522421479225, + 0.0114284111186862, + 0.0242408886551857, + 0.016133036464452744, + 0.008740055374801159, + 0.018003636971116066, + 0.010394389741122723, + 0.008184497244656086, + 0.02187744528055191, + 0.0184704028069973, + 0.023167945444583893, + 0.02449129894375801, + 0.018478183075785637, + 0.016237860545516014, + 0.024266812950372696, + 0.014638680964708328, + 0.01667572557926178, + 0.008453412912786007, + 0.021576611325144768, + 0.007824444212019444, + 0.00786236859858036, + 0.009267653338611126, + 0.006989596877247095, + 0.0066841100342571735, + 0.019909271970391273, + 0.01113958191126585, + 0.02435358613729477, + 0.009310144931077957, + 0.009330075234174728, + 0.013752231374382973, + 0.007927131839096546, + 0.00921577773988247, + 0.0038565690629184246, + 0.004979071673005819, + 0.004608567338436842, + 0.0043229334987699986, + 0.0028610052540898323, + 0.004544798284769058, + 0.00446748873218894, + 0.0034647516440600157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.005978795234113932 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062069907784461975 + ], + "opposite_logprob_tokens": [ + -2.812069892883301 + ], + "logprob_reward": -0.062069907784461975, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062069907784461975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20002402365207672, + 0.14506670832633972, + 0.05843198299407959, + 0.08364344388246536, + 0.14162695407867432, + 0.14557704329490662, + 0.10289008170366287, + 0.040940504521131516, + 0.050510503351688385, + 0.031288791447877884 + ], + "credit_map": [ + 0.0165259949862957, + 0.019877953454852104, + 0.02841625176370144, + 0.038194142282009125, + 0.017595265060663223, + 0.020089665427803993, + 0.035760607570409775, + 0.023564137518405914, + 0.018490510061383247, + 0.018934430554509163, + 0.017136383801698685, + 0.022447412833571434, + 0.010400444269180298, + 0.0143967904150486, + 0.02673567458987236, + 0.016525069251656532, + 0.007067446131259203, + 0.006885141134262085, + 0.006597237195819616, + 0.00890268199145794, + 0.008257976733148098, + 0.004463050048798323, + 0.011868157424032688, + 0.004390287678688765, + 0.011062925681471825, + 0.01096437405794859, + 0.008872686885297298, + 0.012772335670888424, + 0.00861232541501522, + 0.004181873518973589, + 0.02150997705757618, + 0.005666951183229685, + 0.01910293474793434, + 0.01755630224943161, + 0.022780923172831535, + 0.018054969608783722, + 0.011797737330198288, + 0.013975855894386768, + 0.027049031108617783, + 0.011309200897812843, + 0.01734272763133049, + 0.01830173470079899, + 0.01451434101909399, + 0.02238571271300316, + 0.0146641219034791, + 0.020862340927124023, + 0.02389812469482422, + 0.013607939705252647, + 0.01646161638200283, + 0.009393880143761635, + 0.015022533014416695, + 0.015681743621826172, + 0.013625718653202057, + 0.0065018474124372005, + 0.019646238535642624, + 0.006556508131325245, + 0.00556448707357049, + 0.004821580834686756, + 0.007551036775112152, + 0.00418185256421566, + 0.003482674015685916, + 0.006370867602527142, + 0.005147810559719801, + 0.003820196259766817, + 0.006819054018706083, + 0.00639303307980299, + 0.008302854374051094, + 0.007698558736592531, + 0.005901367869228125, + 0.005492174997925758, + 0.004440752789378166, + 0.0054627032950520515, + 0.003817639546468854, + 0.003330615349113941, + 0.0037643439136445522, + 0.004538144916296005, + 0.0032725860364735126, + 0.003302957396954298, + 0.0056265564635396, + 0.0036359489895403385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.062069907784461975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642778541892767 + ], + "opposite_logprob_tokens": [ + -5.379642963409424 + ], + "logprob_reward": -0.004642778541892767, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642778541892767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16654656827449799, + 0.1152353510260582, + 0.04717015475034714, + 0.07516709715127945, + 0.10045994818210602, + 0.13313157856464386, + 0.25778728723526, + 0.04427693039178848, + 0.02709045633673668, + 0.03313454985618591 + ], + "credit_map": [ + 0.016081485897302628, + 0.01857783831655979, + 0.040776558220386505, + 0.03728875145316124, + 0.013889186084270477, + 0.009752945974469185, + 0.019594933837652206, + 0.01058488804847002, + 0.013726339675486088, + 0.017833949998021126, + 0.01113061886280775, + 0.02028006687760353, + 0.008467354811728, + 0.0122254379093647, + 0.02113552950322628, + 0.010436058044433594, + 0.005052455235272646, + 0.004141794051975012, + 0.00963534601032734, + 0.007969985716044903, + 0.004171971697360277, + 0.0030192413832992315, + 0.009478818625211716, + 0.0037005485501140356, + 0.01006107684224844, + 0.010005287826061249, + 0.01216485258191824, + 0.009807482361793518, + 0.006771376356482506, + 0.005238505080342293, + 0.014985957182943821, + 0.0061325631104409695, + 0.015061856247484684, + 0.016039755195379257, + 0.01267822366207838, + 0.013102171011269093, + 0.011029617860913277, + 0.008353116922080517, + 0.016303187236189842, + 0.007892034947872162, + 0.0178880263119936, + 0.01198867242783308, + 0.013007773086428642, + 0.02035113424062729, + 0.013678917661309242, + 0.015584126114845276, + 0.023407617583870888, + 0.017225319519639015, + 0.034968286752700806, + 0.031378261744976044, + 0.02918347343802452, + 0.040181536227464676, + 0.03339216113090515, + 0.021044081076979637, + 0.04289648309350014, + 0.024743037298321724, + 0.007082963362336159, + 0.0050662728026509285, + 0.008275382220745087, + 0.004559435416013002, + 0.006107000634074211, + 0.0030667015817016363, + 0.006104941479861736, + 0.0040142410434782505, + 0.00462058000266552, + 0.003036312758922577, + 0.004798369016498327, + 0.003231003414839506, + 0.0027714294847100973, + 0.002161366632208228, + 0.0035392637364566326, + 0.002932134782895446, + 0.004755717236548662, + 0.0035719245206564665, + 0.005207804962992668, + 0.0038228004705160856, + 0.0040090628899633884, + 0.0038080085068941116, + 0.004004247486591339, + 0.003954990301281214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.004642778541892767 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static macro view of a standard cardboard milk carton on a dark, polished granite countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a tiny, almost imperceptible hole at the bottom of the carton with a single, perfectly opaque white milk droplet expanding from it?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the droplet break free and fall slowly through the still air before impacting the counter surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the droplet shatter into a crown-shaped splash upon impact before collapsing into a small, irregular milky puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a soft bokeh that obscures the kitchen shape while the foreground remains in sharp focus?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007641607895493507 + ], + "opposite_logprob_tokens": [ + -4.882641792297363 + ], + "logprob_reward": -0.007641607895493507, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007641607895493507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20960906147956848, + 0.1278965324163437, + 0.030477134510874748, + 0.1932172328233719, + 0.108455128967762, + 0.08920600265264511, + 0.12519501149654388, + 0.0597265250980854, + 0.03597928583621979, + 0.020238004624843597 + ], + "credit_map": [ + 0.02126879058778286, + 0.015617599710822105, + 0.05488729104399681, + 0.03507288172841072, + 0.010948930867016315, + 0.02466459758579731, + 0.0259318258613348, + 0.021217165514826775, + 0.015841679647564888, + 0.00996292382478714, + 0.0248570553958416, + 0.01757744513452053, + 0.0096623245626688, + 0.01158542837947607, + 0.028796561062335968, + 0.009613130241632462, + 0.0034991158172488213, + 0.003624578006565571, + 0.0031555453315377235, + 0.006827373988926411, + 0.0037244330160319805, + 0.0024116861168295145, + 0.004785492084920406, + 0.0024489141069352627, + 0.02357788011431694, + 0.02835269458591938, + 0.03818013519048691, + 0.03565756604075432, + 0.01520254835486412, + 0.014086863957345486, + 0.02193831093609333, + 0.016221245750784874, + 0.011877063661813736, + 0.013797326944768429, + 0.01993582211434841, + 0.024219509214162827, + 0.009906850755214691, + 0.006890600547194481, + 0.011910500004887581, + 0.009917468763887882, + 0.010818527080118656, + 0.007265071384608746, + 0.020079510286450386, + 0.015443838201463223, + 0.006817423738539219, + 0.00799248181283474, + 0.012217707000672817, + 0.008571453392505646, + 0.019089272245764732, + 0.015614752657711506, + 0.01653491146862507, + 0.02005358226597309, + 0.013977723196148872, + 0.01128536369651556, + 0.018878191709518433, + 0.009761224500834942, + 0.01051167119294405, + 0.006488410755991936, + 0.01229254063218832, + 0.0053695812821388245, + 0.007121091242879629, + 0.006791188847273588, + 0.0056788320653140545, + 0.005473211407661438, + 0.00692326994612813, + 0.0033604097552597523, + 0.007922636345028877, + 0.004604951478540897, + 0.002896183868870139, + 0.0036863803397864103, + 0.0034291488118469715, + 0.0031563090160489082, + 0.0026957537047564983, + 0.0023791994899511337, + 0.003556047100573778, + 0.0027384560089558363, + 0.001446942100301385, + 0.0017256864812225103, + 0.0034649246372282505, + 0.0022309960331767797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.007641607895493507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_1/index_tensor([9475, 9475])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_1/index_tensor([9475, 9475])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2bb4f2278beacfb7ebd13eb0b7daf85c2bf8c106 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_1/index_tensor([9475, 9475])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "majestic high-angle shot triggers a slow, steady downward tilt, revealing a sprawling, hyper-dense metropolis at the peak of the blue hour. Countless skyscrapers, a forest of reflective glass and polished steel, pierce the deep blue sky, their diverse architectural forms creating a complex, richly textured skyline. As the camera continues its deliberate descent, city lights seem to cascade down in a mesmerizing cascade of pure energy. The building facades are an intricate mosaic of thousands of illuminated windows, some radiating the warm, alluring glow of neon, others the cool, sterile white of modern offices, while vibrant, oversized digital billboards splash electric blue and fuchsia across neighboring structures. Far below, the city's arteries throb with activity, the continuous red and white light emitted by vehicular traffic weaving intricate, flowing patterns across deep concrete canyons. The overall atmosphere is one of breathtaking futuristic grandeur, rendered in a cool palette, the deep blue of the night punctuated by vibrant, multi-colored tapestries of artificial light, creating a powerful sense of the city as both a monolithic structure and a living, breathing organism.", + "seed": 5497, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_1/index_9475_seed_5497_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05984950482379645, + "vlm_logprob": -0.05984950482379645, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.05984950482379645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2553510069847107 + ], + "opposite_logprob_tokens": [ + -1.5053510665893555 + ], + "logprob_reward": -0.2553510069847107, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2553510069847107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7386091351509094, + 0.028651800006628036, + 0.022767553105950356, + 0.02116788551211357, + 0.030105987563729286, + 0.01922818459570408, + 0.021145114675164223, + 0.024557681754231453, + 0.045976974070072174, + 0.04778975620865822 + ], + "credit_map": [ + 0.127620130777359, + 0.034685488790273666, + 0.08650512248277664, + 0.035703375935554504, + 0.04657489433884621, + 0.04407816752791405, + 0.26664164662361145, + 0.09680022299289703, + 0.004263672977685928, + 0.004635842051357031, + 0.003262919606640935, + 0.0023938908707350492, + 0.00398417329415679, + 0.003291910281404853, + 0.003305703168734908, + 0.003513684729114175, + 0.002038218080997467, + 0.002070851158350706, + 0.003057321533560753, + 0.0017418785719200969, + 0.0036108584608882666, + 0.0029682933818548918, + 0.0029972658958286047, + 0.004282861948013306, + 0.0030692743603140116, + 0.0018016081303358078, + 0.0031967228278517723, + 0.0028861381579190493, + 0.0020682066679000854, + 0.0032567190937697887, + 0.002726807491853833, + 0.0021624043583869934, + 0.003549004904925823, + 0.0016004134668037295, + 0.005673399195075035, + 0.0020691361278295517, + 0.002529863966628909, + 0.0081405621021986, + 0.0021773537155240774, + 0.004366248846054077, + 0.0019600079394876957, + 0.0010374832199886441, + 0.0031041274778544903, + 0.0023806493263691664, + 0.002461882308125496, + 0.004851156380027533, + 0.0014222166500985622, + 0.002010658150538802, + 0.002123724203556776, + 0.0016441490733996034, + 0.004474259447306395, + 0.0018360684625804424, + 0.004679031204432249, + 0.0031531297136098146, + 0.001659463276155293, + 0.0015752846375107765, + 0.002398504177108407, + 0.0017182865412905812, + 0.0031992807053029537, + 0.003501310246065259, + 0.0043107629753649235, + 0.003085688455030322, + 0.002981462050229311, + 0.003362381597980857, + 0.0051672072149813175, + 0.0013073182199150324, + 0.01442222110927105, + 0.004395366180688143, + 0.002266237512230873, + 0.0034858162980526686, + 0.002861862303689122, + 0.012070935219526291, + 0.0091024124994874, + 0.002600346226245165, + 0.013770087622106075, + 0.005053378641605377, + 0.003452673554420471, + 0.003766357898712158, + 0.003546609077602625, + 0.00649788323789835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.2553510069847107 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370340168476105 + ], + "opposite_logprob_tokens": [ + -3.408703327178955 + ], + "logprob_reward": -0.03370340168476105, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370340168476105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3939869999885559, + 0.12292921543121338, + 0.023243989795446396, + 0.0757107138633728, + 0.07979847490787506, + 0.03373464569449425, + 0.0310995914041996, + 0.018926125019788742, + 0.10606631636619568, + 0.1145038902759552 + ], + "credit_map": [ + 0.06919516623020172, + 0.01938820816576481, + 0.042590755969285965, + 0.022305380553007126, + 0.0257717352360487, + 0.0197516568005085, + 0.14336413145065308, + 0.051619987934827805, + 0.01954217255115509, + 0.02047431841492653, + 0.020209060981869698, + 0.01144938450306654, + 0.008265686221420765, + 0.011909455060958862, + 0.016412580385804176, + 0.014666569419205189, + 0.0028493416029959917, + 0.002399653196334839, + 0.004025634378194809, + 0.0023083011619746685, + 0.002077709650620818, + 0.003274452406913042, + 0.0032657403498888016, + 0.003043159143999219, + 0.00551011273637414, + 0.0027741414960473776, + 0.025666238740086555, + 0.006234750617295504, + 0.003594160545617342, + 0.00779276667162776, + 0.01753685064613819, + 0.006601700559258461, + 0.009198158048093319, + 0.003391817444935441, + 0.021617481485009193, + 0.004591901320964098, + 0.005051910411566496, + 0.014689506031572819, + 0.008613341487944126, + 0.012644371017813683, + 0.0034071602858603, + 0.0024369170423597097, + 0.0041499063372612, + 0.004996071103960276, + 0.00397498020902276, + 0.0054847318679094315, + 0.003426233073696494, + 0.0058586494997143745, + 0.0029990181792527437, + 0.001930702244862914, + 0.007876592688262463, + 0.0027463112492114305, + 0.007307178340852261, + 0.0034323865547776222, + 0.0025933056604117155, + 0.002214098582044244, + 0.0017214748077094555, + 0.0016067824326455593, + 0.0025535051245242357, + 0.002713667694479227, + 0.0026456043124198914, + 0.002513087820261717, + 0.0033292321022599936, + 0.0018427736358717084, + 0.011827847920358181, + 0.0031749876216053963, + 0.031897202134132385, + 0.011874574236571789, + 0.006535568740218878, + 0.009413739666342735, + 0.007203279063105583, + 0.024139124900102615, + 0.01645338162779808, + 0.005747450981289148, + 0.02650097943842411, + 0.012593692168593407, + 0.008784941397607327, + 0.010148069821298122, + 0.01691581681370735, + 0.01735956221818924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03370340168476105 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036140859592705965 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036140859592705965, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036140859592705965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.638772189617157, + 0.13920995593070984, + 0.038876332342624664, + 0.015458988025784492, + 0.030412117019295692, + 0.024749619886279106, + 0.008255726657807827, + 0.013064471073448658, + 0.028320373967289925, + 0.06288027763366699 + ], + "credit_map": [ + 0.1144130602478981, + 0.028833497315645218, + 0.07216200977563858, + 0.029629219323396683, + 0.03851135075092316, + 0.030152589082717896, + 0.24002207815647125, + 0.08504828810691833, + 0.021508019417524338, + 0.019628209993243217, + 0.027338670566678047, + 0.011219383217394352, + 0.008980588987469673, + 0.014587441459298134, + 0.017948364838957787, + 0.017999257892370224, + 0.005356061737984419, + 0.0025840220041573048, + 0.007308095693588257, + 0.003861009841784835, + 0.00366615760140121, + 0.005187036469578743, + 0.006199858151376247, + 0.004714085720479488, + 0.0012350344331935048, + 0.0007746741757728159, + 0.004532902501523495, + 0.0010754381073638797, + 0.0009811593918129802, + 0.0013656888622790575, + 0.004269883967936039, + 0.0012242047814652324, + 0.0032640129793435335, + 0.0011281018378213048, + 0.011491559445858002, + 0.0017713074339553714, + 0.0013954488094896078, + 0.005020417273044586, + 0.0027409358881413937, + 0.0036003312561661005, + 0.002447551814839244, + 0.0017398992786183953, + 0.003660131013020873, + 0.002712428569793701, + 0.0027579935267567635, + 0.0054083820432424545, + 0.002607144182547927, + 0.003416086547076702, + 0.0008968689362518489, + 0.0007205103756859899, + 0.0012569695245474577, + 0.001299519557505846, + 0.0012619305634871125, + 0.0011256139259785414, + 0.000754520995542407, + 0.0009397915564477444, + 0.0013604072155430913, + 0.0008520314004272223, + 0.0018443649169057608, + 0.0019317107507959008, + 0.002256890991702676, + 0.0013945740647614002, + 0.0022941057104617357, + 0.0011303836945444345, + 0.0029230304062366486, + 0.0008393837488256395, + 0.008595635183155537, + 0.0028652995824813843, + 0.0019966773688793182, + 0.002176403533667326, + 0.002224365249276161, + 0.006699577905237675, + 0.011541195213794708, + 0.002608934650197625, + 0.01799871399998665, + 0.0058592078275978565, + 0.004546306561678648, + 0.00512002594769001, + 0.00605775136500597, + 0.009148132987320423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0036140859592705965 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019443194614723325 + ], + "opposite_logprob_tokens": [ + -6.251944541931152 + ], + "logprob_reward": -0.0019443194614723325, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019443194614723325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23338381946086884, + 0.11552323400974274, + 0.11512015014886856, + 0.1421705037355423, + 0.044731125235557556, + 0.06260619312524796, + 0.05370492860674858, + 0.036767806857824326, + 0.024846727028489113, + 0.171145498752594 + ], + "credit_map": [ + 0.03232404589653015, + 0.018435517325997353, + 0.03054533340036869, + 0.02870102971792221, + 0.02252556011080742, + 0.022816209122538567, + 0.04096456989645958, + 0.03707157075405121, + 0.012143812142312527, + 0.008048877120018005, + 0.028146252036094666, + 0.008578542619943619, + 0.00790502317249775, + 0.013679561205208302, + 0.016739625483751297, + 0.020281536504626274, + 0.013129924423992634, + 0.009996902197599411, + 0.015616893768310547, + 0.013548721559345722, + 0.01323967706412077, + 0.01398383267223835, + 0.01814856007695198, + 0.01745564118027687, + 0.013050587847828865, + 0.0070559862069785595, + 0.040132973343133926, + 0.010511419735848904, + 0.00852207001298666, + 0.023749789223074913, + 0.025315433740615845, + 0.013832245022058487, + 0.006527726538479328, + 0.003030280349776149, + 0.009942036122083664, + 0.0024459934793412685, + 0.00296386587433517, + 0.012373964302241802, + 0.0034912927076220512, + 0.003955967258661985, + 0.005034794565290213, + 0.0038744404446333647, + 0.009227372705936432, + 0.006659135688096285, + 0.015149706974625587, + 0.01046939380466938, + 0.005875417031347752, + 0.006315933540463448, + 0.004890934564173222, + 0.003799100173637271, + 0.012393327429890633, + 0.004845202900469303, + 0.01135758776217699, + 0.009260098449885845, + 0.003743149572983384, + 0.003415532410144806, + 0.0034158341586589813, + 0.0021216704044491053, + 0.005447288043797016, + 0.004618550185114145, + 0.008091638796031475, + 0.003637969261035323, + 0.006299213971942663, + 0.0031356406398117542, + 0.002788701094686985, + 0.0018306359415873885, + 0.004389707464724779, + 0.005771785508841276, + 0.002668435452505946, + 0.0027454423252493143, + 0.0025432941038161516, + 0.002108725719153881, + 0.026080746203660965, + 0.008643913082778454, + 0.04637262970209122, + 0.01751687377691269, + 0.013170234858989716, + 0.01704241707921028, + 0.013286296278238297, + 0.029032394289970398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0019443194614723325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634710028767586 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004634710028767586, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634710028767586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2104908972978592, + 0.040838588029146194, + 0.030545739457011223, + 0.05020397901535034, + 0.08843781054019928, + 0.11500753462314606, + 0.06619936972856522, + 0.03241249546408653, + 0.09674356132745743, + 0.2691199481487274 + ], + "credit_map": [ + 0.03703274950385094, + 0.011442163959145546, + 0.022781694307923317, + 0.009901205077767372, + 0.0149879464879632, + 0.011280137114226818, + 0.0761222317814827, + 0.026942776516079903, + 0.005980233661830425, + 0.005448394920676947, + 0.006000599358230829, + 0.004023596178740263, + 0.0027563031762838364, + 0.004943258594721556, + 0.006228677928447723, + 0.0054575237445533276, + 0.0036851793993264437, + 0.0022578961215913296, + 0.005978863686323166, + 0.0029915987979620695, + 0.0031920666806399822, + 0.0038934056647121906, + 0.004514981992542744, + 0.004031747113913298, + 0.004695553332567215, + 0.003235776210203767, + 0.012912126258015633, + 0.003637650515884161, + 0.0031483322381973267, + 0.007080600131303072, + 0.008263057097792625, + 0.007230882067233324, + 0.009399766102433205, + 0.0035565157886594534, + 0.035356197506189346, + 0.006050845608115196, + 0.004402508493512869, + 0.011469917371869087, + 0.008943980559706688, + 0.009258080273866653, + 0.009707793593406677, + 0.007710094563663006, + 0.01960955560207367, + 0.009899437427520752, + 0.024349845945835114, + 0.017923804000020027, + 0.015170988626778126, + 0.010636012069880962, + 0.006294943392276764, + 0.00773418415337801, + 0.009267249144613743, + 0.00764383003115654, + 0.010646180249750614, + 0.012888025492429733, + 0.007107635494321585, + 0.00461732130497694, + 0.0038129668682813644, + 0.002585390117019415, + 0.004255826119333506, + 0.004016906954348087, + 0.004831692203879356, + 0.004991052206605673, + 0.005068514496088028, + 0.002850146032869816, + 0.011293194256722927, + 0.003403260838240385, + 0.02864859439432621, + 0.01403022650629282, + 0.006735495291650295, + 0.006621283013373613, + 0.006253890693187714, + 0.019757622852921486, + 0.033418212085962296, + 0.015042702667415142, + 0.06407462805509567, + 0.025234082713723183, + 0.019895434379577637, + 0.03192511945962906, + 0.024512996897101402, + 0.05501676723361015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.004634710028767586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic high-angle shot triggers a slow, steady downward tilt, revealing a sprawling, hyper-dense metropolis at the peak of the blue hour. Countless skyscrapers, a forest of reflective glass and polished steel, pierce the deep blue sky, their diverse architectural forms creating a complex, richly textured skyline. As the camera continues its deliberate descent, city lights seem to cascade down in a mesmerizing cascade of pure energy. The building facades are an intricate mosaic of thousands of illuminated windows, some radiating the warm, alluring glow of neon, others the cool, sterile white of modern offices, while vibrant, oversized digital billboards splash electric blue and fuchsia across neighboring structures. Far below, the city's arteries throb with activity, the continuous red and white light emitted by vehicular traffic weaving intricate, flowing patterns across deep concrete canyons. The overall atmosphere is one of breathtaking futuristic grandeur, rendered in a cool palette, the deep blue of the night punctuated by vibrant, multi-colored tapestries of artificial light, creating a powerful sense of the city as both a monolithic structure and a living, breathing organism.", + "seed": 5498, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_1/index_9475_seed_5498_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028869859222322703, + "vlm_logprob": -0.028869859222322703, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.028869859222322703, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10246366262435913 + ], + "opposite_logprob_tokens": [ + -2.352463722229004 + ], + "logprob_reward": -0.10246366262435913, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10246366262435913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21978968381881714, + 0.19131547212600708, + 0.05988207086920738, + 0.046457841992378235, + 0.1468687653541565, + 0.11561178416013718, + 0.062100376933813095, + 0.06734837591648102, + 0.041878171265125275, + 0.04874749854207039 + ], + "credit_map": [ + 0.033500246703624725, + 0.02449004165828228, + 0.03006354719400406, + 0.03490247204899788, + 0.025870636105537415, + 0.02561100386083126, + 0.026102188974618912, + 0.01924953982234001, + 0.04074585810303688, + 0.0335441455245018, + 0.011306515894830227, + 0.018815960735082626, + 0.01672104373574257, + 0.01705908216536045, + 0.030021164566278458, + 0.023101676255464554, + 0.008163146674633026, + 0.008007549680769444, + 0.006840708665549755, + 0.010309437289834023, + 0.005409121513366699, + 0.00873477105051279, + 0.006649645045399666, + 0.005767687223851681, + 0.007730028126388788, + 0.004918691702187061, + 0.0038355235010385513, + 0.009665715508162975, + 0.00493395933881402, + 0.006343580782413483, + 0.005145716946572065, + 0.0038846211973577738, + 0.025913244113326073, + 0.034222494810819626, + 0.009929056279361248, + 0.0267543476074934, + 0.015421171672642231, + 0.016222212463617325, + 0.008199592120945454, + 0.01020662859082222, + 0.03716588392853737, + 0.01104323472827673, + 0.005749092437326908, + 0.007620754186064005, + 0.03212582319974899, + 0.01008627563714981, + 0.006251981481909752, + 0.005568724125623703, + 0.007182516157627106, + 0.005431594792753458, + 0.007706476375460625, + 0.007556472439318895, + 0.008113508112728596, + 0.013202067464590073, + 0.0071557071059942245, + 0.005752024240791798, + 0.010129496455192566, + 0.01193914469331503, + 0.005577266216278076, + 0.004234545398503542, + 0.017331959679722786, + 0.007986233569681644, + 0.005728431511670351, + 0.00442128861322999, + 0.004118982702493668, + 0.005351597443223, + 0.005869803950190544, + 0.004391761031001806, + 0.004254207480698824, + 0.005941169802099466, + 0.006809009239077568, + 0.005141634959727526, + 0.004241828806698322, + 0.003506827400997281, + 0.005750066135078669, + 0.009699960239231586, + 0.005534464493393898, + 0.004949377384036779, + 0.008118060417473316, + 0.006946910172700882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.10246366262435913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033708471804857254 + ], + "opposite_logprob_tokens": [ + -3.4087085723876953 + ], + "logprob_reward": -0.033708471804857254, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033708471804857254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17305545508861542, + 0.3034663796424866, + 0.06377578526735306, + 0.06775106489658356, + 0.09133157134056091, + 0.1054927408695221, + 0.06146155297756195, + 0.04856598377227783, + 0.04100010171532631, + 0.04409932345151901 + ], + "credit_map": [ + 0.03947262093424797, + 0.023171406239271164, + 0.02190733328461647, + 0.024113886058330536, + 0.014751814305782318, + 0.01945594698190689, + 0.01710387133061886, + 0.013078578747808933, + 0.057379934936761856, + 0.025384031236171722, + 0.017658241093158722, + 0.025378473103046417, + 0.019256381317973137, + 0.02106381021440029, + 0.09450560808181763, + 0.0428398959338665, + 0.009525753557682037, + 0.006123531609773636, + 0.006373847369104624, + 0.00899474136531353, + 0.007106397300958633, + 0.005569749511778355, + 0.012349740602076054, + 0.007732021156698465, + 0.010387388989329338, + 0.0074066161178052425, + 0.012192091904580593, + 0.006799421738833189, + 0.0062172249890863895, + 0.015068660490214825, + 0.005386755336076021, + 0.004292908124625683, + 0.014111706987023354, + 0.017400868237018585, + 0.005569859873503447, + 0.018725717440247536, + 0.008724324405193329, + 0.014650271274149418, + 0.005930949933826923, + 0.00621787691488862, + 0.029179569333791733, + 0.009829383343458176, + 0.006406846456229687, + 0.007350343279540539, + 0.025942370295524597, + 0.009515318088233471, + 0.007882016710937023, + 0.009386895224452019, + 0.005827729124575853, + 0.0056431288830935955, + 0.006777744274586439, + 0.006032256409525871, + 0.00886701513081789, + 0.013689265586435795, + 0.007018973119556904, + 0.007605442777276039, + 0.0035095270723104477, + 0.008303915150463581, + 0.004644136875867844, + 0.006676285993307829, + 0.006574473809450865, + 0.007261325605213642, + 0.003953356761485338, + 0.0076429639011621475, + 0.0050103869289159775, + 0.004554776009172201, + 0.005384667310863733, + 0.00481448695063591, + 0.005680971313267946, + 0.005422760266810656, + 0.0045879920944571495, + 0.005544060375541449, + 0.005637198220938444, + 0.0035242901649326086, + 0.0053577725775539875, + 0.007565633859485388, + 0.005416858941316605, + 0.0051781050860881805, + 0.00735972635447979, + 0.004059738013893366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.033708471804857254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002488137688487768 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.002488137688487768, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002488137688487768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18339094519615173, + 0.15082880854606628, + 0.04219909757375717, + 0.06174538657069206, + 0.05885881558060646, + 0.3198377192020416, + 0.02718241885304451, + 0.07211843878030777, + 0.03377753123641014, + 0.05006080120801926 + ], + "credit_map": [ + 0.02984275296330452, + 0.027708696201443672, + 0.024962734431028366, + 0.0157928429543972, + 0.016659310087561607, + 0.043807368725538254, + 0.011456371285021305, + 0.0131608871743083, + 0.03158607706427574, + 0.014950323849916458, + 0.01754312589764595, + 0.012924854643642902, + 0.014707133173942566, + 0.02407679706811905, + 0.017436180263757706, + 0.017604341730475426, + 0.005410031881183386, + 0.0045693786814808846, + 0.0036610914394259453, + 0.006976072210818529, + 0.007700329180806875, + 0.0047592525370419025, + 0.004598627332597971, + 0.004524318967014551, + 0.011949622072279453, + 0.007488202303647995, + 0.00500907376408577, + 0.011336731724441051, + 0.006917951628565788, + 0.008070939220488071, + 0.005828646942973137, + 0.005144223105162382, + 0.00923331081867218, + 0.011199519038200378, + 0.004660956095904112, + 0.008074858225882053, + 0.008043684996664524, + 0.008894184604287148, + 0.004529137164354324, + 0.004223170690238476, + 0.09028830379247665, + 0.034078411757946014, + 0.013420243747532368, + 0.024186013266444206, + 0.08964545279741287, + 0.030473493039608, + 0.020117463544011116, + 0.01762833632528782, + 0.002438485622406006, + 0.0026541200932115316, + 0.004057599231600761, + 0.0038999738171696663, + 0.003591762389987707, + 0.004396836739033461, + 0.0025688486639410257, + 0.003574793692678213, + 0.008476383984088898, + 0.008500581607222557, + 0.006935234181582928, + 0.008079132996499538, + 0.012468886561691761, + 0.010977345518767834, + 0.007838532328605652, + 0.008842348121106625, + 0.005125090479850769, + 0.0037815580144524574, + 0.005028833635151386, + 0.004470475949347019, + 0.0036586280912160873, + 0.004504437558352947, + 0.0035580426920205355, + 0.0036504650488495827, + 0.010446002706885338, + 0.0052771721966564655, + 0.0062112994492053986, + 0.004040396306663752, + 0.004940398968756199, + 0.0038516982458531857, + 0.007118011824786663, + 0.00817582942545414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.002488137688487768 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003200886305421591 + ], + "opposite_logprob_tokens": [ + -5.753201007843018 + ], + "logprob_reward": -0.003200886305421591, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003200886305421591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3032667636871338, + 0.1303335577249527, + 0.06871984153985977, + 0.040241789072752, + 0.13584482669830322, + 0.09365098178386688, + 0.06508611887693405, + 0.07329580187797546, + 0.04096414893865585, + 0.04859607666730881 + ], + "credit_map": [ + 0.028880396857857704, + 0.034944236278533936, + 0.04222504422068596, + 0.05442071706056595, + 0.030604427680373192, + 0.03934171050786972, + 0.04251961410045624, + 0.030330635607242584, + 0.026645982638001442, + 0.02453942969441414, + 0.009923421777784824, + 0.01494487002491951, + 0.014833476394414902, + 0.01256611105054617, + 0.01618354022502899, + 0.010696724988520145, + 0.009526080451905727, + 0.009600628167390823, + 0.007072968874126673, + 0.007996970787644386, + 0.007065667305141687, + 0.010453357361257076, + 0.010196736082434654, + 0.006807433906942606, + 0.006287221796810627, + 0.005079925525933504, + 0.003419187618419528, + 0.0068116625770926476, + 0.004920006729662418, + 0.007077098824083805, + 0.0038657214026898146, + 0.002780966693535447, + 0.02359338290989399, + 0.030379170551896095, + 0.006710096728056669, + 0.025481291115283966, + 0.013480989262461662, + 0.01869848184287548, + 0.008695290423929691, + 0.008806130848824978, + 0.01768154464662075, + 0.018027877435088158, + 0.007353753782808781, + 0.012302973307669163, + 0.016910133883357048, + 0.0066018677316606045, + 0.008392146788537502, + 0.0063806818798184395, + 0.00784383062273264, + 0.0054757678881287575, + 0.008051776327192783, + 0.009039671160280704, + 0.008015802130103111, + 0.011515616439282894, + 0.009329007938504219, + 0.005814640782773495, + 0.010616213083267212, + 0.013414500281214714, + 0.005557685624808073, + 0.00353028392419219, + 0.0209569800645113, + 0.009442701935768127, + 0.006549568381160498, + 0.0032278753351420164, + 0.005884963553398848, + 0.004615322221070528, + 0.00626326072961092, + 0.00495119346305728, + 0.004609297960996628, + 0.005647478625178337, + 0.004907514434307814, + 0.004085116554051638, + 0.006993466056883335, + 0.004736536648124456, + 0.005865439306944609, + 0.007072400767356157, + 0.0045080361887812614, + 0.004318799823522568, + 0.00814550556242466, + 0.006955895107239485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.003200886305421591 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002488137688487768 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.002488137688487768, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002488137688487768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1467364877462387, + 0.19307945668697357, + 0.07864764332771301, + 0.04736235737800598, + 0.0739976018667221, + 0.10001789778470993, + 0.1275290995836258, + 0.05954553559422493, + 0.07292353361845016, + 0.1001603975892067 + ], + "credit_map": [ + 0.0183342806994915, + 0.019070953130722046, + 0.017914967611432076, + 0.012823056429624557, + 0.01619824208319187, + 0.042625706642866135, + 0.009358500130474567, + 0.010410778224468231, + 0.026861662045121193, + 0.009513660334050655, + 0.01245040725916624, + 0.01406621653586626, + 0.01481984369456768, + 0.010857371613383293, + 0.07059656828641891, + 0.033913735300302505, + 0.010389404371380806, + 0.008406097069382668, + 0.007071430329233408, + 0.010962322354316711, + 0.00930984877049923, + 0.008473341353237629, + 0.015337402001023293, + 0.008697799406945705, + 0.007640864700078964, + 0.005462228786200285, + 0.00449598440900445, + 0.009577424265444279, + 0.00403548264876008, + 0.008668346330523491, + 0.00432451069355011, + 0.003157515311613679, + 0.013452614657580853, + 0.012663434259593487, + 0.003726702881976962, + 0.011155708692967892, + 0.009165258146822453, + 0.010818205773830414, + 0.00724987406283617, + 0.005765801295638084, + 0.0184974055737257, + 0.017335636541247368, + 0.0072817131876945496, + 0.010349846445024014, + 0.019900906831026077, + 0.010982641950249672, + 0.007864841260015965, + 0.007804908324033022, + 0.013784551061689854, + 0.014174118638038635, + 0.01350751705467701, + 0.011245833709836006, + 0.018117161467671394, + 0.028631163761019707, + 0.013732163235545158, + 0.014336596243083477, + 0.006838420405983925, + 0.007312045898288488, + 0.005279542412608862, + 0.004304998088628054, + 0.013152167201042175, + 0.01054452359676361, + 0.005642102099955082, + 0.006471734493970871, + 0.009537764824926853, + 0.010745454579591751, + 0.0082557313144207, + 0.008843454532325268, + 0.006999076344072819, + 0.0076251341961324215, + 0.011370522901415825, + 0.009546393528580666, + 0.02137298882007599, + 0.010225999169051647, + 0.011540736071765423, + 0.00858489703387022, + 0.01295716967433691, + 0.008154688403010368, + 0.014490162022411823, + 0.01283375546336174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.002488137688487768 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic high-angle shot triggers a slow, steady downward tilt, revealing a sprawling, hyper-dense metropolis at the peak of the blue hour. Countless skyscrapers, a forest of reflective glass and polished steel, pierce the deep blue sky, their diverse architectural forms creating a complex, richly textured skyline. As the camera continues its deliberate descent, city lights seem to cascade down in a mesmerizing cascade of pure energy. The building facades are an intricate mosaic of thousands of illuminated windows, some radiating the warm, alluring glow of neon, others the cool, sterile white of modern offices, while vibrant, oversized digital billboards splash electric blue and fuchsia across neighboring structures. Far below, the city's arteries throb with activity, the continuous red and white light emitted by vehicular traffic weaving intricate, flowing patterns across deep concrete canyons. The overall atmosphere is one of breathtaking futuristic grandeur, rendered in a cool palette, the deep blue of the night punctuated by vibrant, multi-colored tapestries of artificial light, creating a powerful sense of the city as both a monolithic structure and a living, breathing organism.", + "seed": 5499, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_1/index_9475_seed_5499_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017620985861867666, + "vlm_logprob": -0.017620985861867666, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.017620985861867666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06354555487632751 + ], + "opposite_logprob_tokens": [ + -2.8135454654693604 + ], + "logprob_reward": -0.06354555487632751, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06354555487632751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10106170922517776, + 0.10923632234334946, + 0.04482027888298035, + 0.04893507435917854, + 0.03742814436554909, + 0.44469529390335083, + 0.026645051315426826, + 0.04692663997411728, + 0.10851020365953445, + 0.03174129128456116 + ], + "credit_map": [ + 0.0165865458548069, + 0.011539340950548649, + 0.011464802548289299, + 0.014295735396444798, + 0.010448320768773556, + 0.009831320494413376, + 0.008981451392173767, + 0.017914190888404846, + 0.010743976570665836, + 0.009794389829039574, + 0.011547873727977276, + 0.020237194374203682, + 0.02280252054333687, + 0.012431117706000805, + 0.00916462391614914, + 0.012514625675976276, + 0.004786863923072815, + 0.003932253457605839, + 0.0075324163772165775, + 0.004111869260668755, + 0.006511799059808254, + 0.0045799147337675095, + 0.005895447451621294, + 0.007469717413187027, + 0.004631103947758675, + 0.004506705328822136, + 0.005506302695721388, + 0.006618914660066366, + 0.008824801072478294, + 0.005045729223638773, + 0.008040488697588444, + 0.005761025473475456, + 0.0037637220229953527, + 0.003482327563688159, + 0.004643202293664217, + 0.004967596847563982, + 0.004556767176836729, + 0.004416354466229677, + 0.007007282692939043, + 0.004590893629938364, + 0.032202258706092834, + 0.01814677007496357, + 0.00907920766621828, + 0.009756905026733875, + 0.19461075961589813, + 0.1312960833311081, + 0.03392619639635086, + 0.015677105635404587, + 0.0030228535179048777, + 0.004617833532392979, + 0.002540184883400798, + 0.002697654766961932, + 0.003961135167628527, + 0.004612691234797239, + 0.002184521872550249, + 0.003008174244314432, + 0.007325511891394854, + 0.0072411526925861835, + 0.004745165817439556, + 0.006770702078938484, + 0.005858152639120817, + 0.004693418275564909, + 0.004114654380828142, + 0.006177885457873344, + 0.02580403909087181, + 0.01787971518933773, + 0.011950040236115456, + 0.010380887426435947, + 0.012912697158753872, + 0.007488129194825888, + 0.009629359468817711, + 0.012465333566069603, + 0.005477854050695896, + 0.007710602600127459, + 0.002386669861152768, + 0.002273261547088623, + 0.0049667516723275185, + 0.0024324452970176935, + 0.0028522664215415716, + 0.0036414407659322023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 8 + ], + "selected_reward": -0.06354555487632751 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546301819384098 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.012546301819384098, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546301819384098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044195663183927536, + 0.07334821671247482, + 0.019618161022663116, + 0.03527584671974182, + 0.020148025825619698, + 0.7235230803489685, + 0.025221148505806923, + 0.013437317684292793, + 0.03407178819179535, + 0.011160744354128838 + ], + "credit_map": [ + 0.007144542410969734, + 0.004755297210067511, + 0.0049966671504080296, + 0.005766494199633598, + 0.004745117854326963, + 0.004511883947998285, + 0.0038626634050160646, + 0.008413000963628292, + 0.008152004331350327, + 0.005558256525546312, + 0.009745773859322071, + 0.010881271213293076, + 0.012500502169132233, + 0.006092897616326809, + 0.0056997439824044704, + 0.014717776328325272, + 0.0023854398168623447, + 0.002562533365562558, + 0.002926128450781107, + 0.0022828541696071625, + 0.0020516091026365757, + 0.002856830833479762, + 0.0020630143117159605, + 0.0024897530674934387, + 0.002539003500714898, + 0.00451411260291934, + 0.0036689320113509893, + 0.005983741022646427, + 0.005947949830442667, + 0.0033862211275845766, + 0.00564958481118083, + 0.0035863055381923914, + 0.0025904965586960316, + 0.0021621370688080788, + 0.0021734817419201136, + 0.0024538743309676647, + 0.002502366667613387, + 0.0034609613940119743, + 0.0021532676182687283, + 0.002651442075148225, + 0.05439792200922966, + 0.030966240912675858, + 0.01646711677312851, + 0.015751786530017853, + 0.319978803396225, + 0.20290978252887726, + 0.05579926073551178, + 0.027252180501818657, + 0.0024189879186451435, + 0.0038828470278531313, + 0.0014858818612992764, + 0.0047958712093532085, + 0.00316175096668303, + 0.004070310853421688, + 0.0023082986008375883, + 0.0030972023960202932, + 0.002398960990831256, + 0.0030348184518516064, + 0.0011721556074917316, + 0.00116952252574265, + 0.0019270364427939057, + 0.0013027915265411139, + 0.0010731677757576108, + 0.0013588662259280682, + 0.007324536330997944, + 0.004786534234881401, + 0.0033172217663377523, + 0.0033932714723050594, + 0.003589117433875799, + 0.002351833041757345, + 0.004695116076618433, + 0.004614159930497408, + 0.0013450158294290304, + 0.0017755632288753986, + 0.0009650388965383172, + 0.0009037402342073619, + 0.001479469588957727, + 0.0013025592779740691, + 0.0017750216647982597, + 0.0016143360408023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.012546301819384098 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011781663633883 + ], + "opposite_logprob_tokens": [ + -6.75117826461792 + ], + "logprob_reward": -0.0011781663633883, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011781663633883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07165035605430603, + 0.10915286839008331, + 0.13778595626354218, + 0.031722500920295715, + 0.039361849427223206, + 0.5029559135437012, + 0.028557591140270233, + 0.02401500567793846, + 0.025901230052113533, + 0.028896713629364967 + ], + "credit_map": [ + 0.01203970704227686, + 0.007748288568109274, + 0.009470023214817047, + 0.006982266437262297, + 0.006312723271548748, + 0.006826826371252537, + 0.007625246420502663, + 0.014645268209278584, + 0.010043364018201828, + 0.00958738662302494, + 0.013366997241973877, + 0.018800929188728333, + 0.022119438275694847, + 0.01041197869926691, + 0.01221161987632513, + 0.012611158192157745, + 0.018641183152794838, + 0.011563573032617569, + 0.021377336233854294, + 0.011448885314166546, + 0.015411173924803734, + 0.013303712010383606, + 0.012890471145510674, + 0.03314962238073349, + 0.003002840792760253, + 0.006587756332010031, + 0.0036619058810174465, + 0.0023022745735943317, + 0.005270210560411215, + 0.003765292465686798, + 0.003240911290049553, + 0.0038913071621209383, + 0.004990942310541868, + 0.004861138761043549, + 0.0038558889646083117, + 0.004969516769051552, + 0.005495085380971432, + 0.005339328665286303, + 0.004101383499801159, + 0.00574856624007225, + 0.03608287125825882, + 0.021733669564127922, + 0.010676868259906769, + 0.011371826753020287, + 0.21899031102657318, + 0.1470370590686798, + 0.0378851555287838, + 0.019178146496415138, + 0.003386352676898241, + 0.0053265150636434555, + 0.001872019493021071, + 0.003476640908047557, + 0.004381412174552679, + 0.005522960796952248, + 0.0021170885302126408, + 0.002474600449204445, + 0.003416038816794753, + 0.0044455439783632755, + 0.0016661740373820066, + 0.0030364750418812037, + 0.003211570903658867, + 0.002244591945782304, + 0.0025294197257608175, + 0.003465192625299096, + 0.00445985421538353, + 0.0040071625262498856, + 0.0032232399098575115, + 0.0025352907832711935, + 0.0029961385298520327, + 0.0021117227151989937, + 0.002706069964915514, + 0.0038617511745542288, + 0.0023382699582725763, + 0.004788931459188461, + 0.002137679373845458, + 0.0034457992296665907, + 0.003692853730171919, + 0.003824631217867136, + 0.005009188316762447, + 0.0036593603435903788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.0011781663633883 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004102861508727074 + ], + "opposite_logprob_tokens": [ + -5.50410270690918 + ], + "logprob_reward": -0.004102861508727074, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004102861508727074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38542312383651733, + 0.08655471354722977, + 0.09202750027179718, + 0.09216009825468063, + 0.033356793224811554, + 0.16083842515945435, + 0.05239640548825264, + 0.0279006939381361, + 0.0428333580493927, + 0.026508929207921028 + ], + "credit_map": [ + 0.05967298150062561, + 0.035730063915252686, + 0.04080522432923317, + 0.04861720651388168, + 0.06015069782733917, + 0.04661504179239273, + 0.042101118713617325, + 0.05173076316714287, + 0.008198804222047329, + 0.007535192649811506, + 0.01063371729105711, + 0.015035273507237434, + 0.0192136000841856, + 0.009371462278068066, + 0.00813366286456585, + 0.008433003909885883, + 0.009569012559950352, + 0.013425798155367374, + 0.011822836473584175, + 0.01143380906432867, + 0.0093063460662961, + 0.01432839035987854, + 0.010872638784348965, + 0.011268671602010727, + 0.008588983677327633, + 0.008091175928711891, + 0.009851344861090183, + 0.01307118684053421, + 0.018292374908924103, + 0.009244464337825775, + 0.014650664292275906, + 0.010369904339313507, + 0.0031016243156045675, + 0.0031286878511309624, + 0.004764650482684374, + 0.003404814749956131, + 0.0060242172330617905, + 0.0035893043968826532, + 0.004981541074812412, + 0.004361950792372227, + 0.011272047646343708, + 0.006233133375644684, + 0.004314800258725882, + 0.005031890701502562, + 0.06388489902019501, + 0.05284777283668518, + 0.01163503061980009, + 0.005618846043944359, + 0.005185624584555626, + 0.006808715872466564, + 0.0037248400039970875, + 0.0074124401435256, + 0.007717642467468977, + 0.008368616923689842, + 0.0060087232850492, + 0.0071698022074997425, + 0.0041467780247330666, + 0.005433395970612764, + 0.002805085387080908, + 0.003350208280608058, + 0.003956437576562166, + 0.002614399651065469, + 0.0025090775452554226, + 0.0030853115022182465, + 0.008558068424463272, + 0.006568979471921921, + 0.004885097965598106, + 0.0040621026419103146, + 0.004699182230979204, + 0.0038239683490246534, + 0.004396018572151661, + 0.005839942023158073, + 0.00278536812402308, + 0.003556777723133564, + 0.0023260980378836393, + 0.0024208417162299156, + 0.003236263059079647, + 0.0032477856148034334, + 0.004258768167346716, + 0.004677026532590389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.004102861508727074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732044741511345 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006732044741511345, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732044741511345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08666099607944489, + 0.05404157564043999, + 0.04926653206348419, + 0.03743944317102432, + 0.09854952991008759, + 0.45643478631973267, + 0.02946208231151104, + 0.058945342898368835, + 0.05319598689675331, + 0.07600381225347519 + ], + "credit_map": [ + 0.011426117271184921, + 0.006876196246594191, + 0.012891273945569992, + 0.013725357130169868, + 0.008614081889390945, + 0.010695372708141804, + 0.008087076246738434, + 0.014345511794090271, + 0.006589704658836126, + 0.004127549938857555, + 0.006739443633705378, + 0.007121228612959385, + 0.009822100400924683, + 0.004140071105211973, + 0.005173784680664539, + 0.010327688418328762, + 0.0053780036978423595, + 0.0056141456589102745, + 0.006385662127286196, + 0.006353418342769146, + 0.006804532837122679, + 0.005805617664009333, + 0.00581083120778203, + 0.007114320062100887, + 0.004223319701850414, + 0.0048749069683253765, + 0.004482128657400608, + 0.0036288381088525057, + 0.004642346408218145, + 0.004679265432059765, + 0.004286250099539757, + 0.006622387561947107, + 0.010973773896694183, + 0.009687377139925957, + 0.009131059981882572, + 0.008383804000914097, + 0.016950858756899834, + 0.016199568286538124, + 0.013485245406627655, + 0.01373783778399229, + 0.031853314489126205, + 0.020176365971565247, + 0.009358171373605728, + 0.0114896260201931, + 0.1937735676765442, + 0.13920941948890686, + 0.03443721681833267, + 0.01613708771765232, + 0.0036211838014423847, + 0.003889520186930895, + 0.0029136966913938522, + 0.0039011016488075256, + 0.004798553418368101, + 0.004040455911308527, + 0.002424811013042927, + 0.003872754517942667, + 0.01311502605676651, + 0.013665466569364071, + 0.004054916091263294, + 0.004683360457420349, + 0.008598222397267818, + 0.0058537074364721775, + 0.004424654878675938, + 0.004549980629235506, + 0.00641204696148634, + 0.00747646251693368, + 0.005725734401494265, + 0.006475300993770361, + 0.005360019393265247, + 0.005198252387344837, + 0.0075331493280828, + 0.009015018120408058, + 0.006503999698907137, + 0.011637684889137745, + 0.005519240163266659, + 0.010341032408177853, + 0.009160101413726807, + 0.013590359129011631, + 0.009268086403608322, + 0.009983302094042301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.006732044741511345 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic high-angle shot triggers a slow, steady downward tilt, revealing a sprawling, hyper-dense metropolis at the peak of the blue hour. Countless skyscrapers, a forest of reflective glass and polished steel, pierce the deep blue sky, their diverse architectural forms creating a complex, richly textured skyline. As the camera continues its deliberate descent, city lights seem to cascade down in a mesmerizing cascade of pure energy. The building facades are an intricate mosaic of thousands of illuminated windows, some radiating the warm, alluring glow of neon, others the cool, sterile white of modern offices, while vibrant, oversized digital billboards splash electric blue and fuchsia across neighboring structures. Far below, the city's arteries throb with activity, the continuous red and white light emitted by vehicular traffic weaving intricate, flowing patterns across deep concrete canyons. The overall atmosphere is one of breathtaking futuristic grandeur, rendered in a cool palette, the deep blue of the night punctuated by vibrant, multi-colored tapestries of artificial light, creating a powerful sense of the city as both a monolithic structure and a living, breathing organism.", + "seed": 5500, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_1/index_9475_seed_5500_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07348603191785515, + "vlm_logprob": -0.07348603191785515, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.07348603191785515, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3162602186203003 + ], + "opposite_logprob_tokens": [ + -1.3162602186203003 + ], + "logprob_reward": -0.3162602186203003, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3162602186203003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06888551265001297, + 0.037891026586294174, + 0.016232870519161224, + 0.035429295152425766, + 0.024982580915093422, + 0.024746373295783997, + 0.7326934337615967, + 0.016223149374127388, + 0.020641835406422615, + 0.022273950278759003 + ], + "credit_map": [ + 0.009951231069862843, + 0.007360814139246941, + 0.006083356682211161, + 0.005515946075320244, + 0.009801405481994152, + 0.016316838562488556, + 0.00680135190486908, + 0.007054570131003857, + 0.005306541919708252, + 0.006769937928766012, + 0.0033486599568277597, + 0.003472344484180212, + 0.006056812591850758, + 0.004565664567053318, + 0.004477330483496189, + 0.003893738379701972, + 0.0028714111540466547, + 0.002010696567595005, + 0.001703606452792883, + 0.002207130193710327, + 0.0020535546354949474, + 0.0014966171002015471, + 0.002005450427532196, + 0.0018844048026949167, + 0.005999830551445484, + 0.005856236908584833, + 0.004215670749545097, + 0.0062172627076506615, + 0.0038198421243578196, + 0.0036698984913527966, + 0.0032796834129840136, + 0.0023708739317953587, + 0.0025734526570886374, + 0.002655884949490428, + 0.00253432709723711, + 0.0025403271429240704, + 0.004438355099409819, + 0.0033636216539889574, + 0.003556439420208335, + 0.0033201759215444326, + 0.0026937678921967745, + 0.003021717770025134, + 0.0029288530349731445, + 0.003988403361290693, + 0.0029562211129814386, + 0.003869894426316023, + 0.0025923047214746475, + 0.0026952119078487158, + 0.08437852561473846, + 0.028956441208720207, + 0.2504541873931885, + 0.06740936636924744, + 0.05925997719168663, + 0.19134275615215302, + 0.01701144501566887, + 0.03388078883290291, + 0.00310531840659678, + 0.0020163790322840214, + 0.002528771525248885, + 0.002419213531538844, + 0.001442549517378211, + 0.0014562889700755477, + 0.00170118676032871, + 0.0015534424455836415, + 0.0028447641525417566, + 0.002189712133258581, + 0.003365549258887768, + 0.0029509030282497406, + 0.0032211453653872013, + 0.0018097949214279652, + 0.0024388006422668695, + 0.001821167883463204, + 0.0029885561671108007, + 0.002473108470439911, + 0.0036339638754725456, + 0.0022810003720223904, + 0.003014969639480114, + 0.002204790711402893, + 0.003378491150215268, + 0.0022990733850747347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.3162602186203003 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306967929005623 + ], + "opposite_logprob_tokens": [ + -3.16806960105896 + ], + "logprob_reward": -0.04306967929005623, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306967929005623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05773574486374855, + 0.031103387475013733, + 0.016961660236120224, + 0.019789887592196465, + 0.0232552383095026, + 0.023628082126379013, + 0.7155801653862, + 0.0651123970746994, + 0.0353362113237381, + 0.011497207917273045 + ], + "credit_map": [ + 0.008611333556473255, + 0.0056622447445988655, + 0.006878591142594814, + 0.006819566246122122, + 0.006073968484997749, + 0.012353979982435703, + 0.005213591270148754, + 0.006122466642409563, + 0.0030196188017725945, + 0.006894102320075035, + 0.0027199473697692156, + 0.002386998850852251, + 0.002383950399234891, + 0.0017825884278863668, + 0.008583751507103443, + 0.003332429565489292, + 0.0033863636199384928, + 0.0019953050650656223, + 0.0022179398220032454, + 0.00214863708242774, + 0.0020032802131026983, + 0.001680129673331976, + 0.001892084488645196, + 0.001637920388020575, + 0.0021283281967043877, + 0.001951287966221571, + 0.0022924658842384815, + 0.003454037941992283, + 0.002542360918596387, + 0.002940334612503648, + 0.0030006037559360266, + 0.0014804687816649675, + 0.00266086682677269, + 0.0026808041147887707, + 0.0026079146191477776, + 0.003285219892859459, + 0.0039517758414149284, + 0.0030703728552907705, + 0.00296584889292717, + 0.002032434567809105, + 0.002295092446729541, + 0.0027156269643455744, + 0.002896926598623395, + 0.0034077269956469536, + 0.004868017043918371, + 0.0022872372064739466, + 0.002812018385156989, + 0.002345435554161668, + 0.08329544216394424, + 0.028407501056790352, + 0.24364206194877625, + 0.0656481459736824, + 0.05831700190901756, + 0.18660202622413635, + 0.016768774017691612, + 0.03289920836687088, + 0.009603066369891167, + 0.009496664628386497, + 0.011637300252914429, + 0.009648463688790798, + 0.005575462244451046, + 0.005737628787755966, + 0.006091244984418154, + 0.007322569377720356, + 0.0036452030763030052, + 0.002847908530384302, + 0.005494655109941959, + 0.00458762189373374, + 0.003345165867358446, + 0.0038259709253907204, + 0.0047842529602348804, + 0.006805431563407183, + 0.001708462368696928, + 0.0008529025944881141, + 0.0017886054702103138, + 0.0014600929571315646, + 0.0009483215981163085, + 0.0006912420503795147, + 0.002123800339177251, + 0.0019237813539803028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.04306967929005623 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015124559868127108 + ], + "opposite_logprob_tokens": [ + -6.50151252746582 + ], + "logprob_reward": -0.0015124559868127108, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015124559868127108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08662685751914978, + 0.38118401169776917, + 0.02108040265738964, + 0.07367269694805145, + 0.06910071521997452, + 0.10523875057697296, + 0.20240634679794312, + 0.014468585141003132, + 0.015155677683651447, + 0.031065892428159714 + ], + "credit_map": [ + 0.008851924911141396, + 0.014145242981612682, + 0.0073660751804709435, + 0.009263184852898121, + 0.01670168898999691, + 0.0101358937099576, + 0.011474309489130974, + 0.0086885467171669, + 0.03041379339993, + 0.10176251083612442, + 0.02584921382367611, + 0.02930428460240364, + 0.03080655261874199, + 0.02435600757598877, + 0.11226484179496765, + 0.02642681635916233, + 0.0029987511225044727, + 0.0023491738829761744, + 0.0021900231949985027, + 0.0036410395987331867, + 0.002654947340488434, + 0.0026094962377101183, + 0.002051041228696704, + 0.0025859312154352665, + 0.007330428808927536, + 0.010135813616216183, + 0.011222421191632748, + 0.01536750327795744, + 0.006641326937824488, + 0.007138151675462723, + 0.010099157691001892, + 0.005737903527915478, + 0.009778935462236404, + 0.008287430740892887, + 0.00879665371030569, + 0.005790728144347668, + 0.015143102034926414, + 0.009781388565897942, + 0.006821244955062866, + 0.004701236728578806, + 0.00886167399585247, + 0.01217996422201395, + 0.010560611262917519, + 0.01239217072725296, + 0.024986138567328453, + 0.019772501662373543, + 0.010852670297026634, + 0.005633035209029913, + 0.0247085839509964, + 0.01558294240385294, + 0.06198994442820549, + 0.02095903642475605, + 0.019076112657785416, + 0.0344650074839592, + 0.013617157004773617, + 0.012007578276097775, + 0.0032286138739436865, + 0.0013383803889155388, + 0.0020242026075720787, + 0.00236661359667778, + 0.0019455980509519577, + 0.0011328374966979027, + 0.0013687661848962307, + 0.0010635757353156805, + 0.0016686072340235114, + 0.001706863404251635, + 0.002505786484107375, + 0.0033107250928878784, + 0.0018030490027740598, + 0.0010931919096037745, + 0.002090142108500004, + 0.0009773133788257837, + 0.00300887250341475, + 0.0025780333671718836, + 0.006848583929240704, + 0.003985071554780006, + 0.0026290540117770433, + 0.002807445591315627, + 0.005535914562642574, + 0.0036729208659380674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.0015124559868127108 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024952725507318974 + ], + "opposite_logprob_tokens": [ + -6.002495288848877 + ], + "logprob_reward": -0.0024952725507318974, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024952725507318974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18601511418819427, + 0.08442667871713638, + 0.11665285378694534, + 0.05775527283549309, + 0.045576319098472595, + 0.05045367404818535, + 0.32896852493286133, + 0.035903241485357285, + 0.04993360862135887, + 0.044314734637737274 + ], + "credit_map": [ + 0.028280923143029213, + 0.018549593165516853, + 0.01808164268732071, + 0.021632466465234756, + 0.022905850782990456, + 0.039090365171432495, + 0.019455626606941223, + 0.018018649891018867, + 0.006900912616401911, + 0.008770500309765339, + 0.006081438157707453, + 0.007375519722700119, + 0.015789251774549484, + 0.010159144178032875, + 0.01846289075911045, + 0.010887020267546177, + 0.013022924773395061, + 0.016337862238287926, + 0.013613414950668812, + 0.01855078712105751, + 0.012221155688166618, + 0.009742538444697857, + 0.011929928325116634, + 0.021234238520264626, + 0.0070720333606004715, + 0.010911649093031883, + 0.009585355408489704, + 0.0059313480742275715, + 0.005869534332305193, + 0.0056954026222229, + 0.007698332890868187, + 0.004991614725440741, + 0.004791441839188337, + 0.005458031315356493, + 0.0049340794794261456, + 0.008745475672185421, + 0.005152123514562845, + 0.005592061672359705, + 0.004387813154608011, + 0.006515292916446924, + 0.005445363465696573, + 0.008852124214172363, + 0.005606299266219139, + 0.005076075438410044, + 0.008685259148478508, + 0.008206751197576523, + 0.004682882223278284, + 0.0038989183958619833, + 0.03843865543603897, + 0.01469317264854908, + 0.11473485082387924, + 0.031257618218660355, + 0.02466917596757412, + 0.08072283118963242, + 0.008762910962104797, + 0.015689266845583916, + 0.01041255984455347, + 0.0030527410563081503, + 0.0057217092253267765, + 0.004143216647207737, + 0.0033366968855261803, + 0.0028641237877309322, + 0.002838582033291459, + 0.0035336092114448547, + 0.00659399526193738, + 0.0041737849824130535, + 0.00898632314056158, + 0.008171225897967815, + 0.005891002248972654, + 0.004004310350865126, + 0.008089378476142883, + 0.004023585002869368, + 0.007466815412044525, + 0.0046737054362893105, + 0.004952108021825552, + 0.0037820579018443823, + 0.0051912181079387665, + 0.004833986982703209, + 0.007961302064359188, + 0.005453538615256548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.0024952725507318974 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a sprawling, hyper-dense metropolis with countless skyscrapers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle shot and perform a slow, steady downward tilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the time of day depicted as the blue hour with a deep blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the building facades shown as reflective glass and polished steel with illuminated windows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there vibrant, oversized digital billboards displaying electric blue and fuchsia colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092533141374588 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092533141374588, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092533141374588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09371469169855118, + 0.046718478202819824, + 0.20286612212657928, + 0.1185639500617981, + 0.08986899256706238, + 0.08575725555419922, + 0.18354809284210205, + 0.06578399240970612, + 0.03565520793199539, + 0.07752326130867004 + ], + "credit_map": [ + 0.014158250764012337, + 0.009968996047973633, + 0.015732644125819206, + 0.011622858233749866, + 0.009226896800100803, + 0.012528207153081894, + 0.012068495154380798, + 0.008408339694142342, + 0.006147645879536867, + 0.01125468872487545, + 0.003948659170418978, + 0.00434007728472352, + 0.005335337016731501, + 0.004743590019643307, + 0.006681120488792658, + 0.004267358686774969, + 0.023317039012908936, + 0.025526829063892365, + 0.022875050082802773, + 0.03035675175487995, + 0.02361038140952587, + 0.017384644597768784, + 0.02108636125922203, + 0.03870907425880432, + 0.009625144302845001, + 0.011909105814993382, + 0.015476875938475132, + 0.023601027205586433, + 0.011663340032100677, + 0.01683652214705944, + 0.016278982162475586, + 0.013172955252230167, + 0.013351736590266228, + 0.009494463913142681, + 0.008939665742218494, + 0.0105271702632308, + 0.02020140178501606, + 0.012701524421572685, + 0.008593511767685413, + 0.0060595120303332806, + 0.008925164118409157, + 0.009995785541832447, + 0.009363085962831974, + 0.013417634181678295, + 0.011433190666139126, + 0.014878321439027786, + 0.01005215011537075, + 0.007691919803619385, + 0.02172692120075226, + 0.01484724972397089, + 0.050204500555992126, + 0.02291768603026867, + 0.017985651269555092, + 0.02784014120697975, + 0.01571284793317318, + 0.012313097715377808, + 0.01660454086959362, + 0.00642093550413847, + 0.010760648176074028, + 0.0089110741391778, + 0.007795227225869894, + 0.0048032780177891254, + 0.005507833324372768, + 0.004980454687029123, + 0.004273558035492897, + 0.0037661814130842686, + 0.004853236023336649, + 0.0051466622389853, + 0.0037641115486621857, + 0.0035803995560854673, + 0.006160921882838011, + 0.004110137932002544, + 0.008068091236054897, + 0.008334202691912651, + 0.012372162193059921, + 0.008540933020412922, + 0.00806251261383295, + 0.008864650502800941, + 0.012387446127831936, + 0.010893267579376698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.004092533141374588 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_10/index_tensor([152, 152])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_10/index_tensor([152, 152])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..052b29d991dd9868a9019e73eb06df9612e5f79d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_10/index_tensor([152, 152])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "single, static, high-angle, extreme-wide shot captures a catastrophic event, rendered entirely in Picasso's bold, fragmented, and multi-perspective style. A powerful mountain torrent surges across the landscape, not a flowing liquid but a turbulent collision of sharp, geometric planes in ochre browns and deep amber. This destructive force carves a path through the dense jungle of Java water cress, its delicate, parsley-like leaves deconstructed into jagged, acidic green shapes that churn and shatter within the Cubist vortex of roiling brown water. The scene is viewed simultaneously from multiple, impossible angles; we simultaneously see the flat surface of the floodwater, the cross-section of the water tearing through the soil, and the outlines of the waves. Against a dark, oppressive sky composed of scarred gray and black polygons, the raw power and destructive geometry of the event take center stage, creating a chaotic, flattened yet vibrant composition that emphasizes form and violent movement over realistic depiction.", + "seed": 11484, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_10/index_152_seed_11484_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.4553343251347542, + "vlm_logprob": -0.4553343251347542, + "vlm_yes_no_margin": 1.225, + "vlm_token_credit": -0.4553343251347542, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0544646978378296 + ], + "opposite_logprob_tokens": [ + -0.429464727640152 + ], + "logprob_reward": -1.0544646978378296, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0544646978378296, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09652683883905411, + 0.10076470673084259, + 0.06781411916017532, + 0.045438092201948166, + 0.01943524368107319, + 0.04829423129558563, + 0.020673565566539764, + 0.39435699582099915, + 0.09773112088441849, + 0.10896513611078262 + ], + "credit_map": [ + 0.007099547889083624, + 0.01584952510893345, + 0.012168646790087223, + 0.013527841307222843, + 0.0115509619936347, + 0.008551550097763538, + 0.010266775265336037, + 0.017511984333395958, + 0.010488249361515045, + 0.010918372310698032, + 0.012131303548812866, + 0.012371736578643322, + 0.013861749321222305, + 0.013185270130634308, + 0.016899386420845985, + 0.010908621363341808, + 0.006091349758207798, + 0.010157343000173569, + 0.00971620436757803, + 0.009401116520166397, + 0.006342724896967411, + 0.005254972726106644, + 0.010886259377002716, + 0.009964140132069588, + 0.0031494845170527697, + 0.004325715359300375, + 0.004129220265895128, + 0.008930621668696404, + 0.003987075295299292, + 0.004738913848996162, + 0.006990109570324421, + 0.009186946786940098, + 0.002330738352611661, + 0.002572649624198675, + 0.0020544796716421843, + 0.00268455664627254, + 0.002549270633608103, + 0.0020047782454639673, + 0.002311395714059472, + 0.002927373629063368, + 0.003855734597891569, + 0.004454373382031918, + 0.006273957435041666, + 0.006406231317669153, + 0.0058930483646690845, + 0.003917028661817312, + 0.010546460747718811, + 0.006947391666471958, + 0.004369309172034264, + 0.001962409121915698, + 0.001526852953247726, + 0.0030451214406639338, + 0.0017180551076307893, + 0.003526965156197548, + 0.0022630011662840843, + 0.002261849818751216, + 0.02900869958102703, + 0.03747916594147682, + 0.0395800806581974, + 0.066086046397686, + 0.05641970783472061, + 0.03485463932156563, + 0.06811711937189102, + 0.06281150877475739, + 0.006692955736070871, + 0.005940418690443039, + 0.022181442007422447, + 0.02929224818944931, + 0.00605468126013875, + 0.00368124688975513, + 0.007449683267623186, + 0.01643843948841095, + 0.021439094096422195, + 0.0091429203748703, + 0.016537103801965714, + 0.01315388921648264, + 0.008659390732645988, + 0.011753634549677372, + 0.015603095293045044, + 0.012676003389060497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 1 + ], + "selected_reward": -1.0544646978378296 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004737854003906 + ], + "opposite_logprob_tokens": [ + -2.695047378540039 + ], + "logprob_reward": -0.07004737854003906, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004737854003906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0718001052737236, + 0.04737172648310661, + 0.053614288568496704, + 0.036877237260341644, + 0.03701206296682358, + 0.07069480419158936, + 0.02124616503715515, + 0.41939184069633484, + 0.13046422600746155, + 0.11152749508619308 + ], + "credit_map": [ + 0.006377263460308313, + 0.01392288040369749, + 0.006970182526856661, + 0.011538905091583729, + 0.00782797485589981, + 0.006929092109203339, + 0.006107033230364323, + 0.012126780115067959, + 0.005748870316892862, + 0.004827569704502821, + 0.005200191866606474, + 0.003337037982419133, + 0.006953308824449778, + 0.008668946102261543, + 0.006296020001173019, + 0.00633978471159935, + 0.004799344576895237, + 0.004704515915364027, + 0.01082486193627119, + 0.006576957646757364, + 0.006684667430818081, + 0.0035048488061875105, + 0.0081781642511487, + 0.008340928703546524, + 0.0034261513501405716, + 0.005255797412246466, + 0.0050425538793206215, + 0.005477194674313068, + 0.003216282930225134, + 0.0032726062927395105, + 0.00578099861741066, + 0.0054056537337601185, + 0.00353868561796844, + 0.004796653985977173, + 0.0034287916496396065, + 0.00407103868201375, + 0.0037549780681729317, + 0.004601058084517717, + 0.008137024939060211, + 0.00468383077532053, + 0.005677279084920883, + 0.0059177870862185955, + 0.00961080938577652, + 0.01086390856653452, + 0.0063925571739673615, + 0.004852086305618286, + 0.017257532104849815, + 0.010122844018042088, + 0.005860405042767525, + 0.002336029661819339, + 0.0020180705469101667, + 0.0028429199010133743, + 0.0016977295745164156, + 0.0024055002722889185, + 0.0020620061550289392, + 0.0020235024858266115, + 0.01659253053367138, + 0.01761980913579464, + 0.027019793167710304, + 0.0339859314262867, + 0.03345451131463051, + 0.02012861706316471, + 0.13785916566848755, + 0.13273151218891144, + 0.0088739562779665, + 0.007148813921958208, + 0.02912248857319355, + 0.04048178717494011, + 0.007696179207414389, + 0.004402145277708769, + 0.009473505429923534, + 0.023265346884727478, + 0.019712377339601517, + 0.009467109106481075, + 0.014934073202311993, + 0.015219809487462044, + 0.009382170625030994, + 0.012154333293437958, + 0.01790340058505535, + 0.012754210270941257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.07004737854003906 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029817529022693634 + ], + "opposite_logprob_tokens": [ + -3.529817581176758 + ], + "logprob_reward": -0.029817529022693634, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029817529022693634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07777079939842224, + 0.053058117628097534, + 0.19125260412693024, + 0.0749361664056778, + 0.04527147859334946, + 0.041442956775426865, + 0.019604068249464035, + 0.36593806743621826, + 0.0645689070224762, + 0.06615686416625977 + ], + "credit_map": [ + 0.006076314952224493, + 0.015096462331712246, + 0.008525742217898369, + 0.010904396884143353, + 0.009363860823214054, + 0.009102649986743927, + 0.008680950850248337, + 0.010020417161285877, + 0.005274087190628052, + 0.01030823029577732, + 0.006025777198374271, + 0.0057946620509028435, + 0.00560805993154645, + 0.005624196957796812, + 0.007506712339818478, + 0.006916389800608158, + 0.019140638411045074, + 0.0330137275159359, + 0.028836041688919067, + 0.01884619891643524, + 0.029783012345433235, + 0.01367625780403614, + 0.02724127657711506, + 0.020715437829494476, + 0.003439839230850339, + 0.004221830051392317, + 0.005359048955142498, + 0.02324233204126358, + 0.0040847924537956715, + 0.0045695556327700615, + 0.007104756776243448, + 0.022914007306098938, + 0.004332955926656723, + 0.003944568336009979, + 0.005167027935385704, + 0.008158281445503235, + 0.0035344725474715233, + 0.003805591957643628, + 0.01072035450488329, + 0.005608225706964731, + 0.003123093396425247, + 0.0039198207668960094, + 0.00432912353426218, + 0.0056826770305633545, + 0.004584517329931259, + 0.0035772293340414762, + 0.009753222577273846, + 0.006473270244896412, + 0.003563967999070883, + 0.0022786781191825867, + 0.0015865565510466695, + 0.002604587934911251, + 0.0017575272358953953, + 0.0029746948275715113, + 0.001933220075443387, + 0.0029048335272818804, + 0.0265289768576622, + 0.032928574830293655, + 0.031426627188920975, + 0.047912005335092545, + 0.04557602480053902, + 0.030075639486312866, + 0.06809388101100922, + 0.0833963081240654, + 0.004745940212160349, + 0.0038898929487913847, + 0.01494920440018177, + 0.0189115721732378, + 0.004380212631076574, + 0.0026347823441028595, + 0.00535843288525939, + 0.009698864072561264, + 0.00854115467518568, + 0.00562145933508873, + 0.01420807745307684, + 0.01012581866234541, + 0.004093711264431477, + 0.0064544291235506535, + 0.009775551967322826, + 0.0073366533033549786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.029817529022693634 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934325098991394 + ], + "opposite_logprob_tokens": [ + -0.6934325098991394 + ], + "logprob_reward": -0.6934325098991394, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934325098991394, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15311099588871002, + 0.08871598541736603, + 0.061073530465364456, + 0.06734127551317215, + 0.07106824964284897, + 0.041456807404756546, + 0.03865208476781845, + 0.292864054441452, + 0.08177638053894043, + 0.10394058376550674 + ], + "credit_map": [ + 0.015425856225192547, + 0.03612586110830307, + 0.014054112136363983, + 0.016224723309278488, + 0.01651364006102085, + 0.01765279471874237, + 0.016098951920866966, + 0.021015066653490067, + 0.010008560493588448, + 0.012872937135398388, + 0.007296557538211346, + 0.010635839775204659, + 0.013722577132284641, + 0.011101942509412766, + 0.011514903977513313, + 0.011562670581042767, + 0.004791141487658024, + 0.010264136828482151, + 0.011179128661751747, + 0.006225212477147579, + 0.009474185295403004, + 0.004174753092229366, + 0.008256131783127785, + 0.006708841305226088, + 0.0040779756382107735, + 0.0042065586894750595, + 0.011031730100512505, + 0.014187196269631386, + 0.00629848288372159, + 0.006239232607185841, + 0.00876576080918312, + 0.012534335255622864, + 0.00514057045802474, + 0.0064690858125686646, + 0.006265978794544935, + 0.008456695824861526, + 0.00655531557276845, + 0.009332624264061451, + 0.018858501687645912, + 0.009989477694034576, + 0.003470180556178093, + 0.005843050312250853, + 0.005863355938345194, + 0.006306063383817673, + 0.0050426507368683815, + 0.0036065164022147655, + 0.0060112872160971165, + 0.005313703324645758, + 0.003996205050498247, + 0.005063140299171209, + 0.003883902681991458, + 0.006231962703168392, + 0.004209712613373995, + 0.005380838643759489, + 0.004083550069481134, + 0.0058027757331728935, + 0.016357602551579475, + 0.026407308876514435, + 0.019668668508529663, + 0.027884969487786293, + 0.04047790542244911, + 0.02121271751821041, + 0.07935560494661331, + 0.06149929016828537, + 0.006377921439707279, + 0.0056824046187102795, + 0.020064234733581543, + 0.022614048793911934, + 0.005655315238982439, + 0.0038635707460343838, + 0.006459275260567665, + 0.01105961762368679, + 0.011531872674822807, + 0.009958806447684765, + 0.022595645859837532, + 0.013335958123207092, + 0.007466581184417009, + 0.010376465506851673, + 0.01630658283829689, + 0.012368669733405113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.6934325098991394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289095103740692 + ], + "opposite_logprob_tokens": [ + -1.0539095401763916 + ], + "logprob_reward": -0.4289095103740692, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289095103740692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07075612992048264, + 0.04178447276353836, + 0.10402492433786392, + 0.039769869297742844, + 0.020822573453187943, + 0.05989206209778786, + 0.015292974188923836, + 0.44175589084625244, + 0.1550491601228714, + 0.05085194483399391 + ], + "credit_map": [ + 0.006332548335194588, + 0.013997051864862442, + 0.0068241567350924015, + 0.007719364948570728, + 0.007656961213797331, + 0.009323404170572758, + 0.008333045057952404, + 0.010569596663117409, + 0.005430177319794893, + 0.003475509351119399, + 0.003935567568987608, + 0.005509865004569292, + 0.0053604478016495705, + 0.00675071170553565, + 0.0055902269668877125, + 0.005731966812163591, + 0.010967371053993702, + 0.01895291917026043, + 0.015250945463776588, + 0.009915710426867008, + 0.016083888709545135, + 0.008185986429452896, + 0.013542347587645054, + 0.011125758290290833, + 0.003131493693217635, + 0.0036359180230647326, + 0.005571534391492605, + 0.008256293833255768, + 0.0027638543397188187, + 0.003638423979282379, + 0.006108314264565706, + 0.006664038635790348, + 0.00222049024887383, + 0.0023178965784609318, + 0.0024734577164053917, + 0.002940990962088108, + 0.002245294163003564, + 0.00199867389164865, + 0.003987638279795647, + 0.0026381309144198895, + 0.0047965217381715775, + 0.004963231738656759, + 0.007479981053620577, + 0.009329511784017086, + 0.0054436069913208485, + 0.004623868502676487, + 0.014378429390490055, + 0.008876912295818329, + 0.0018564987694844604, + 0.0016107206465676427, + 0.0013945873361080885, + 0.002649382222443819, + 0.0017669516382738948, + 0.002183940727263689, + 0.002002763096243143, + 0.0018281291704624891, + 0.02505493350327015, + 0.03221912682056427, + 0.03291803225874901, + 0.04890130087733269, + 0.04952487722039223, + 0.027555853128433228, + 0.12687565386295319, + 0.09870611131191254, + 0.010913144797086716, + 0.008973491378128529, + 0.035537928342819214, + 0.0464860163629055, + 0.010032995603978634, + 0.005540370009839535, + 0.01133453194051981, + 0.026230689138174057, + 0.008063380606472492, + 0.0054174927063286304, + 0.00820230133831501, + 0.006288574542850256, + 0.003899655072018504, + 0.0053823357447981834, + 0.0067725712433457375, + 0.006825633812695742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.4289095103740692 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "single, static, high-angle, extreme-wide shot captures a catastrophic event, rendered entirely in Picasso's bold, fragmented, and multi-perspective style. A powerful mountain torrent surges across the landscape, not a flowing liquid but a turbulent collision of sharp, geometric planes in ochre browns and deep amber. This destructive force carves a path through the dense jungle of Java water cress, its delicate, parsley-like leaves deconstructed into jagged, acidic green shapes that churn and shatter within the Cubist vortex of roiling brown water. The scene is viewed simultaneously from multiple, impossible angles; we simultaneously see the flat surface of the floodwater, the cross-section of the water tearing through the soil, and the outlines of the waves. Against a dark, oppressive sky composed of scarred gray and black polygons, the raw power and destructive geometry of the event take center stage, creating a chaotic, flattened yet vibrant composition that emphasizes form and violent movement over realistic depiction.", + "seed": 11485, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_10/index_152_seed_11485_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14359565936028956, + "vlm_logprob": -0.14359565936028956, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.14359565936028956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22648212313652039 + ], + "opposite_logprob_tokens": [ + -1.6014821529388428 + ], + "logprob_reward": -0.22648212313652039, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22648212313652039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20470745861530304, + 0.06817954033613205, + 0.0755457878112793, + 0.2243521511554718, + 0.07538158446550369, + 0.11611481755971909, + 0.045043136924505234, + 0.06126384064555168, + 0.06828145682811737, + 0.06113025173544884 + ], + "credit_map": [ + 0.024793753400444984, + 0.02638060599565506, + 0.02086416818201542, + 0.017169209197163582, + 0.044173527508974075, + 0.03252720087766647, + 0.018720680847764015, + 0.020078297704458237, + 0.009063177742064, + 0.006000729743391275, + 0.009586862288415432, + 0.013834606856107712, + 0.007964138872921467, + 0.00792750809341669, + 0.008330672979354858, + 0.005471833981573582, + 0.008839004673063755, + 0.010276323184370995, + 0.008420654572546482, + 0.006497609429061413, + 0.008285529911518097, + 0.015742920339107513, + 0.006388969719409943, + 0.011094771325588226, + 0.014488879591226578, + 0.029950400814414024, + 0.009654943831264973, + 0.009960615076124668, + 0.07979299873113632, + 0.010388534516096115, + 0.011313265189528465, + 0.05880249664187431, + 0.01196298561990261, + 0.008671264164149761, + 0.011000980623066425, + 0.007130212616175413, + 0.010596317239105701, + 0.007959822192788124, + 0.008402944542467594, + 0.009657046757638454, + 0.008764971978962421, + 0.011390658095479012, + 0.006599638145416975, + 0.004258239641785622, + 0.03378802165389061, + 0.038142260164022446, + 0.00765020539984107, + 0.005520810838788748, + 0.0033686780370771885, + 0.0028063494246453047, + 0.012936597689986229, + 0.0022478036116808653, + 0.0034470476675778627, + 0.015909964218735695, + 0.001488685142248869, + 0.0028380074072629213, + 0.00781832356005907, + 0.0071181138046085835, + 0.012300523929297924, + 0.010268821381032467, + 0.008352337405085564, + 0.005560509394854307, + 0.004188287537544966, + 0.005656914319843054, + 0.006964921019971371, + 0.010567797347903252, + 0.00922383088618517, + 0.0063536870293319225, + 0.010311898775398731, + 0.0059754992835223675, + 0.009473713114857674, + 0.009410105645656586, + 0.010135505348443985, + 0.0095071317628026, + 0.006306719500571489, + 0.006743360310792923, + 0.007016527466475964, + 0.007924964651465416, + 0.007094391621649265, + 0.00640165014192462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.22648212313652039 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079863533377647 + ], + "opposite_logprob_tokens": [ + -4.141079902648926 + ], + "logprob_reward": -0.016079863533377647, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079863533377647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18363088369369507, + 0.050590790808200836, + 0.16114871203899384, + 0.18694248795509338, + 0.07585780322551727, + 0.06050052121281624, + 0.1143348440527916, + 0.06770522147417068, + 0.0683690533041954, + 0.030919663608074188 + ], + "credit_map": [ + 0.020689360797405243, + 0.02466105856001377, + 0.02180125005543232, + 0.017171218991279602, + 0.04278746619820595, + 0.0240788534283638, + 0.016516154631972313, + 0.01592552475631237, + 0.0075199659913778305, + 0.010919985361397266, + 0.007284540217369795, + 0.004979128483682871, + 0.004967986140400171, + 0.005065914709120989, + 0.004330846946686506, + 0.00552242249250412, + 0.018228068947792053, + 0.02697780542075634, + 0.018672926351428032, + 0.016207240521907806, + 0.02603723667562008, + 0.026632746681571007, + 0.013301312923431396, + 0.015091367065906525, + 0.012222354300320148, + 0.014079432934522629, + 0.014993250370025635, + 0.01087737176567316, + 0.0544840432703495, + 0.008932956494390965, + 0.018599586561322212, + 0.05275348573923111, + 0.012498260475695133, + 0.006983151193708181, + 0.012118321843445301, + 0.006984810344874859, + 0.009984620846807957, + 0.008123893290758133, + 0.007731180172413588, + 0.011433565057814121, + 0.009051576256752014, + 0.008200261741876602, + 0.0055541666224598885, + 0.0036080467980355024, + 0.015897618606686592, + 0.00746339512988925, + 0.005521764513105154, + 0.005203689448535442, + 0.007872398942708969, + 0.007025284226983786, + 0.03596939146518707, + 0.004768291022628546, + 0.006127546075731516, + 0.04483257606625557, + 0.002415180904790759, + 0.005324173253029585, + 0.007349597755819559, + 0.00812822300940752, + 0.012376180849969387, + 0.011901584453880787, + 0.007278921082615852, + 0.0051551638171076775, + 0.00695325993001461, + 0.008562286384403706, + 0.008699179627001286, + 0.007333396468311548, + 0.012334417551755905, + 0.005502895452082157, + 0.01076543889939785, + 0.005835722666233778, + 0.010226824320852757, + 0.0076711755245924, + 0.004799195099622011, + 0.003912863787263632, + 0.0033023704309016466, + 0.003050844417884946, + 0.004279996734112501, + 0.003714627120643854, + 0.004086399916559458, + 0.003773365868255496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.016079863533377647 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370801359415054 + ], + "opposite_logprob_tokens": [ + -3.408708095550537 + ], + "logprob_reward": -0.03370801359415054, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370801359415054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1980675905942917, + 0.03314187377691269, + 0.06433966755867004, + 0.048795707523822784, + 0.05585050210356712, + 0.08550703525543213, + 0.27756646275520325, + 0.10702972859144211, + 0.09726367145776749, + 0.03243774548172951 + ], + "credit_map": [ + 0.03588133305311203, + 0.020826293155550957, + 0.01158444955945015, + 0.010887461714446545, + 0.039316434413194656, + 0.05763821676373482, + 0.010321492329239845, + 0.011611915193498135, + 0.004581385292112827, + 0.006099980790168047, + 0.0035468486603349447, + 0.002720359480008483, + 0.00629798648878932, + 0.004463124554604292, + 0.002334744203835726, + 0.003097445936873555, + 0.007183159701526165, + 0.009687448851764202, + 0.010087481699883938, + 0.0065500433556735516, + 0.007995550520718098, + 0.01030086912214756, + 0.006377245299518108, + 0.006157866679131985, + 0.005478208884596825, + 0.004673677962273359, + 0.004785099066793919, + 0.00788678415119648, + 0.00968511588871479, + 0.004047227092087269, + 0.005741700995713472, + 0.0064978948794305325, + 0.008755337446928024, + 0.0064567201770842075, + 0.004923912230879068, + 0.0033416368532925844, + 0.0075829606503248215, + 0.010451700538396835, + 0.006226848345249891, + 0.008111384697258472, + 0.0076862722635269165, + 0.010505642741918564, + 0.005493296775966883, + 0.003240948310121894, + 0.02059994637966156, + 0.02333017624914646, + 0.010501677170395851, + 0.004149072337895632, + 0.018641989678144455, + 0.015479423105716705, + 0.08797238022089005, + 0.011898691765964031, + 0.015307864174246788, + 0.10859394818544388, + 0.006045331247150898, + 0.01362682692706585, + 0.013003489002585411, + 0.015794849023222923, + 0.016615407541394234, + 0.012944183312356472, + 0.010463633574545383, + 0.01638401672244072, + 0.009879153221845627, + 0.011945001780986786, + 0.008907421492040157, + 0.01078417431563139, + 0.014538413845002651, + 0.007953332737088203, + 0.015672150999307632, + 0.007617948576807976, + 0.016536980867385864, + 0.015253251418471336, + 0.00499756308272481, + 0.003907025326043367, + 0.003714409191161394, + 0.004320585168898106, + 0.004161481745541096, + 0.0040792254731059074, + 0.003869064152240753, + 0.0033883897121995687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.03370801359415054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813628315925598 + ], + "opposite_logprob_tokens": [ + -1.406362771987915 + ], + "logprob_reward": -0.2813628315925598, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813628315925598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17381246387958527, + 0.04277985170483589, + 0.05993729457259178, + 0.13257528841495514, + 0.06393181532621384, + 0.04957736283540726, + 0.34981459379196167, + 0.046410754323005676, + 0.053555794060230255, + 0.027604833245277405 + ], + "credit_map": [ + 0.026439806446433067, + 0.02861318178474903, + 0.01707957126200199, + 0.01819317787885666, + 0.026217572391033173, + 0.019730838015675545, + 0.018770569935441017, + 0.018767723813652992, + 0.007007652428001165, + 0.008536077104508877, + 0.007184356916695833, + 0.0035450190771371126, + 0.0047187055461108685, + 0.003843608545139432, + 0.0036031389608979225, + 0.00434128986671567, + 0.00773911876603961, + 0.00814177468419075, + 0.009940974414348602, + 0.005729207769036293, + 0.006871028803288937, + 0.007970733568072319, + 0.005342963617295027, + 0.008201487362384796, + 0.012900947593152523, + 0.03787452355027199, + 0.007520240731537342, + 0.014138774015009403, + 0.02305111289024353, + 0.011645175516605377, + 0.013109263963997364, + 0.012335242703557014, + 0.010842346586287022, + 0.004793220199644566, + 0.00899093970656395, + 0.005716107785701752, + 0.008819092065095901, + 0.008562122471630573, + 0.006074612494558096, + 0.010133369825780392, + 0.0029428291600197554, + 0.0035204654559493065, + 0.0030411917250603437, + 0.0019490455742925406, + 0.016814621165394783, + 0.015488898381590843, + 0.003140658838674426, + 0.0026796425227075815, + 0.023490121588110924, + 0.019713303074240685, + 0.11353336274623871, + 0.01502438448369503, + 0.019821038469672203, + 0.13344474136829376, + 0.0070710377767682076, + 0.01771659031510353, + 0.006771139334887266, + 0.005915957968682051, + 0.006916412152349949, + 0.005373185034841299, + 0.006773192901164293, + 0.0064323777332901955, + 0.0033989360090345144, + 0.004829547367990017, + 0.008082461543381214, + 0.013942432589828968, + 0.006045990157872438, + 0.003639789531007409, + 0.004025633912533522, + 0.0050138384103775024, + 0.005421306472271681, + 0.0073843346908688545, + 0.004349156282842159, + 0.004369776230305433, + 0.004022272303700447, + 0.0028516435995697975, + 0.003196962643414736, + 0.003091940190643072, + 0.0031184805557131767, + 0.002604598645120859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.2813628315925598 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034546494483948 + ], + "opposite_logprob_tokens": [ + -1.910345435142517 + ], + "logprob_reward": -0.16034546494483948, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034546494483948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3718750476837158, + 0.06015573814511299, + 0.07818327099084854, + 0.13077032566070557, + 0.13778284192085266, + 0.01820995658636093, + 0.08425701409578323, + 0.032664958387613297, + 0.06153807416558266, + 0.024562768638134003 + ], + "credit_map": [ + 0.025456218048930168, + 0.050666384398937225, + 0.019153442233800888, + 0.031200502067804337, + 0.1247606873512268, + 0.08661860972642899, + 0.016515066847205162, + 0.017504137009382248, + 0.008290871046483517, + 0.006751049775630236, + 0.0076562510803341866, + 0.00993617158383131, + 0.008840647526085377, + 0.006495858076959848, + 0.005968607496470213, + 0.006216281093657017, + 0.008349004201591015, + 0.017520809546113014, + 0.0073938979767262936, + 0.008286328986287117, + 0.009212358854711056, + 0.013235487975180149, + 0.005766877438873053, + 0.00841850508004427, + 0.010564750991761684, + 0.012470449320971966, + 0.00733190169557929, + 0.009720329195261002, + 0.03830328956246376, + 0.008935324847698212, + 0.009374297223985195, + 0.03406998887658119, + 0.01945260912179947, + 0.010938924737274647, + 0.02073606289923191, + 0.012251327745616436, + 0.013538317754864693, + 0.0227888822555542, + 0.011669817380607128, + 0.026406895369291306, + 0.002277220832183957, + 0.0030729316640645266, + 0.001576596056111157, + 0.0015274531906470656, + 0.0025195146445184946, + 0.0029052288737148046, + 0.0023038508370518684, + 0.0020271597895771265, + 0.0059699080884456635, + 0.007136065978556871, + 0.024027269333600998, + 0.0037958482280373573, + 0.004826322663575411, + 0.03190118446946144, + 0.001916371169500053, + 0.004684039391577244, + 0.003429598407819867, + 0.004565967246890068, + 0.0034818609710782766, + 0.004007148090749979, + 0.003239911049604416, + 0.006660218816250563, + 0.0028411168605089188, + 0.00443913834169507, + 0.0069261835888028145, + 0.01956099458038807, + 0.004926260560750961, + 0.0032800964545458555, + 0.012082160450518131, + 0.004524391144514084, + 0.005670864600688219, + 0.004567120689898729, + 0.0029936833307147026, + 0.003163591492921114, + 0.0025281300768256187, + 0.0035701203159987926, + 0.002972034038975835, + 0.003147448878735304, + 0.0028985324315726757, + 0.003289229469373822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.16034546494483948 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, high-angle, extreme-wide shot captures a catastrophic event, rendered entirely in Picasso's bold, fragmented, and multi-perspective style. A powerful mountain torrent surges across the landscape, not a flowing liquid but a turbulent collision of sharp, geometric planes in ochre browns and deep amber. This destructive force carves a path through the dense jungle of Java water cress, its delicate, parsley-like leaves deconstructed into jagged, acidic green shapes that churn and shatter within the Cubist vortex of roiling brown water. The scene is viewed simultaneously from multiple, impossible angles; we simultaneously see the flat surface of the floodwater, the cross-section of the water tearing through the soil, and the outlines of the waves. Against a dark, oppressive sky composed of scarred gray and black polygons, the raw power and destructive geometry of the event take center stage, creating a chaotic, flattened yet vibrant composition that emphasizes form and violent movement over realistic depiction.", + "seed": 11486, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_10/index_152_seed_11486_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03828350938856602, + "vlm_logprob": -0.03828350938856602, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.03828350938856602, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019004639238119125 + ], + "opposite_logprob_tokens": [ + -4.019004821777344 + ], + "logprob_reward": -0.019004639238119125, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019004639238119125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3230176270008087, + 0.12280961126089096, + 0.058689557015895844, + 0.10026317834854126, + 0.1282569319009781, + 0.028607139363884926, + 0.06903121620416641, + 0.08583679050207138, + 0.023275334388017654, + 0.06021261215209961 + ], + "credit_map": [ + 0.03150135651230812, + 0.017357291653752327, + 0.06884004920721054, + 0.028546055778861046, + 0.020807502791285515, + 0.021657098084688187, + 0.09010918438434601, + 0.04419908672571182, + 0.008803820237517357, + 0.0066617559641599655, + 0.011191453784704208, + 0.017535174265503883, + 0.007186656817793846, + 0.00881593581289053, + 0.029645567759871483, + 0.03296924754977226, + 0.009296976961195469, + 0.003663631621748209, + 0.006598809268325567, + 0.010973688215017319, + 0.004774243570864201, + 0.005104451440274715, + 0.008169358596205711, + 0.010108395479619503, + 0.009800509549677372, + 0.01021764799952507, + 0.012787151150405407, + 0.026077866554260254, + 0.00851892028003931, + 0.007491307333111763, + 0.014560938812792301, + 0.010808838531374931, + 0.016792230308055878, + 0.012327313423156738, + 0.01605033501982689, + 0.01579342968761921, + 0.01403223630040884, + 0.017349224537611008, + 0.017452310770750046, + 0.01845984347164631, + 0.0029328702948987484, + 0.0030861604027450085, + 0.005399018060415983, + 0.005589371547102928, + 0.002442619763314724, + 0.0026302747428417206, + 0.0034750306513160467, + 0.003051792737096548, + 0.00546086672693491, + 0.004382949322462082, + 0.008661765605211258, + 0.015586297027766705, + 0.00508091039955616, + 0.004657942336052656, + 0.014092792756855488, + 0.011107692494988441, + 0.0064514088444411755, + 0.0052330889739096165, + 0.016177376732230186, + 0.018247950822114944, + 0.006302191875874996, + 0.006952264811843634, + 0.01515422947704792, + 0.011318281292915344, + 0.003779710503295064, + 0.0016322560841217637, + 0.0035513106267899275, + 0.0031514489091932774, + 0.0016243952559307218, + 0.0019662941340357065, + 0.003351410385221243, + 0.004218508489429951, + 0.004032706841826439, + 0.004165238700807095, + 0.0035576012451201677, + 0.004812057130038738, + 0.004715300630778074, + 0.008158493787050247, + 0.016964580863714218, + 0.013806630857288837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.019004639238119125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898162305355072 + ], + "opposite_logprob_tokens": [ + -2.578981637954712 + ], + "logprob_reward": -0.07898162305355072, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898162305355072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17421771585941315, + 0.2664979100227356, + 0.046549342572689056, + 0.08305617421865463, + 0.23430395126342773, + 0.01573297195136547, + 0.06955377757549286, + 0.04416521266102791, + 0.021273421123623848, + 0.044649478048086166 + ], + "credit_map": [ + 0.021224478259682655, + 0.014010860584676266, + 0.021165186539292336, + 0.018383225426077843, + 0.01096821203827858, + 0.011957373470067978, + 0.05074954032897949, + 0.02575884759426117, + 0.0163026861846447, + 0.013124085031449795, + 0.0199868306517601, + 0.036464136093854904, + 0.010873313061892986, + 0.012073516845703125, + 0.0891646221280098, + 0.06850871443748474, + 0.005333180073648691, + 0.0032536638900637627, + 0.006787607911974192, + 0.01090693473815918, + 0.0031401279848068953, + 0.004041329026222229, + 0.007258296944200993, + 0.005828200839459896, + 0.006197616457939148, + 0.0070034959353506565, + 0.01202573161572218, + 0.022295963019132614, + 0.006620486732572317, + 0.006961836479604244, + 0.010783419944345951, + 0.011167623102664948, + 0.023536166176199913, + 0.015982192009687424, + 0.02631736733019352, + 0.042209722101688385, + 0.018250828608870506, + 0.020603584125638008, + 0.04870664328336716, + 0.03869744762778282, + 0.0016802032478153706, + 0.0019043459324166179, + 0.0021827600430697203, + 0.0024110074155032635, + 0.001530469162389636, + 0.0017928497400134802, + 0.0021692521404474974, + 0.002062082989141345, + 0.004617522470653057, + 0.0026273648254573345, + 0.022774890065193176, + 0.010793671011924744, + 0.0036250248085707426, + 0.005338113289326429, + 0.006446473766118288, + 0.013330718502402306, + 0.0042414590716362, + 0.0034763908479362726, + 0.007019737735390663, + 0.008642766624689102, + 0.002865083049982786, + 0.0036815269850194454, + 0.009067943319678307, + 0.0051703061908483505, + 0.002350202528759837, + 0.0018124368507415056, + 0.0030291182920336723, + 0.004220162518322468, + 0.0014062192058190703, + 0.0015874684322625399, + 0.002834220649674535, + 0.00403359392657876, + 0.0013848121743649244, + 0.001303663244470954, + 0.002104670275002718, + 0.003868126543238759, + 0.0018911749357357621, + 0.002352588577196002, + 0.01748819276690483, + 0.014256251975893974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.07898162305355072 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059071511030197 + ], + "opposite_logprob_tokens": [ + -3.8955907821655273 + ], + "logprob_reward": -0.02059071511030197, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059071511030197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09578600525856018, + 0.1015653908252716, + 0.029121719300746918, + 0.03694735839962959, + 0.27823516726493835, + 0.014432178810238838, + 0.14332793653011322, + 0.04141756147146225, + 0.0258050374686718, + 0.2333616316318512 + ], + "credit_map": [ + 0.00954015739262104, + 0.004166456870734692, + 0.008883742615580559, + 0.008745837025344372, + 0.00462618563324213, + 0.007165260147303343, + 0.02709406241774559, + 0.025564298033714294, + 0.00773116759955883, + 0.006551262456923723, + 0.010504364036023617, + 0.017940936610102654, + 0.005228234454989433, + 0.005700814072042704, + 0.020017368718981743, + 0.02789122611284256, + 0.0035316634457558393, + 0.0022142655216157436, + 0.0032176454551517963, + 0.007117193192243576, + 0.001967688323929906, + 0.001897444948554039, + 0.005957597866654396, + 0.003218218684196472, + 0.003079501213505864, + 0.002341704210266471, + 0.006528541445732117, + 0.010727055370807648, + 0.002629956929013133, + 0.0024984334595501423, + 0.004440344404429197, + 0.004701816942542791, + 0.026731576770544052, + 0.021092379465699196, + 0.033651310950517654, + 0.046642910689115524, + 0.02362126111984253, + 0.03055102936923504, + 0.057273250073194504, + 0.03867141529917717, + 0.0022582218516618013, + 0.0011492919875308871, + 0.0020619381684809923, + 0.001866228529252112, + 0.0013812227407470345, + 0.0017508777091279626, + 0.001966578885912895, + 0.0019978168420493603, + 0.007244454696774483, + 0.006963479332625866, + 0.03571672365069389, + 0.03317319229245186, + 0.006314762867987156, + 0.00859485287219286, + 0.027689380571246147, + 0.01763107255101204, + 0.0033329527359455824, + 0.002693998161703348, + 0.006303166504949331, + 0.00876547023653984, + 0.002774760127067566, + 0.0029560260009020567, + 0.008254854939877987, + 0.0063363295048475266, + 0.002256255131214857, + 0.0010766031919047236, + 0.004456217400729656, + 0.005792014300823212, + 0.001443158951587975, + 0.0019049979746341705, + 0.004247377160936594, + 0.004628411028534174, + 0.0063804155215620995, + 0.002690310589969158, + 0.009070189669728279, + 0.01559015829116106, + 0.008328357711434364, + 0.008217650465667248, + 0.09472924470901489, + 0.08835529536008835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.02059071511030197 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787445440888405 + ], + "opposite_logprob_tokens": [ + -3.529787540435791 + ], + "logprob_reward": -0.029787445440888405, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787445440888405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13831204175949097, + 0.17604608833789825, + 0.04593454301357269, + 0.03858619183301926, + 0.3310270309448242, + 0.020383493974804878, + 0.043352190405130386, + 0.03515741229057312, + 0.019486548379063606, + 0.15171438455581665 + ], + "credit_map": [ + 0.01587018370628357, + 0.008244624361395836, + 0.03579455241560936, + 0.015934133902192116, + 0.010146502405405045, + 0.009625432081520557, + 0.016739418730139732, + 0.025957221165299416, + 0.007906182669103146, + 0.010566456243395805, + 0.013060683384537697, + 0.029185323044657707, + 0.00815585721284151, + 0.009775059297680855, + 0.03823935613036156, + 0.059157177805900574, + 0.004665078595280647, + 0.004228951875120401, + 0.004409922752529383, + 0.011277077719569206, + 0.0038162111304700375, + 0.0027112409006804228, + 0.008546405471861362, + 0.006279663182795048, + 0.0041704862378537655, + 0.004513015504926443, + 0.0058319722302258015, + 0.0064597162418067455, + 0.004028763622045517, + 0.0043288348242640495, + 0.004334473982453346, + 0.004918933380395174, + 0.03144470229744911, + 0.025713039562106133, + 0.038022883236408234, + 0.05804486572742462, + 0.028140192851424217, + 0.0334874764084816, + 0.0683491975069046, + 0.04782472550868988, + 0.002711045788601041, + 0.0019088218687102199, + 0.003192664822563529, + 0.003019987605512142, + 0.0016611414030194283, + 0.0025391827803105116, + 0.0023740739561617374, + 0.0029765774961560965, + 0.003839621553197503, + 0.004049376118928194, + 0.011918275617063046, + 0.005125961732119322, + 0.00338996690697968, + 0.004037049133330584, + 0.0044443681836128235, + 0.006547575816512108, + 0.003239390905946493, + 0.004470371175557375, + 0.003980971407145262, + 0.0039962660521268845, + 0.005469478201121092, + 0.004165357444435358, + 0.0044589778408408165, + 0.005376603454351425, + 0.0026246155612170696, + 0.001845506951212883, + 0.0022966249380260706, + 0.0038317753933370113, + 0.0016284642042592168, + 0.001991644036024809, + 0.0023827163968235254, + 0.0028852024115622044, + 0.004334415774792433, + 0.0029400254134088755, + 0.006081665400415659, + 0.010267642326653004, + 0.005477888509631157, + 0.006006447598338127, + 0.061176832765340805, + 0.05542948842048645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.029787445440888405 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043053124099969864 + ], + "opposite_logprob_tokens": [ + -3.168053150177002 + ], + "logprob_reward": -0.043053124099969864, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043053124099969864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1680571287870407, + 0.07033489644527435, + 0.046554114669561386, + 0.04997697472572327, + 0.20097112655639648, + 0.029637370258569717, + 0.03799569606781006, + 0.01891528069972992, + 0.022191757336258888, + 0.3553655445575714 + ], + "credit_map": [ + 0.013665597885847092, + 0.006406786851584911, + 0.01229478232562542, + 0.013936679810285568, + 0.007607536390423775, + 0.010334910824894905, + 0.05251019820570946, + 0.05130065977573395, + 0.00418239226564765, + 0.004584425128996372, + 0.006388893350958824, + 0.01739429496228695, + 0.0038147768937051296, + 0.004160368349403143, + 0.007999553345143795, + 0.021810192614793777, + 0.004068442154675722, + 0.004586250986903906, + 0.00505999056622386, + 0.01020070631057024, + 0.0045288666151463985, + 0.0036700402852147818, + 0.007776952348649502, + 0.006662867497652769, + 0.0024824305437505245, + 0.002681637881323695, + 0.008990505710244179, + 0.017169229686260223, + 0.0022534711752086878, + 0.002916408935561776, + 0.0049682194367051125, + 0.008515078574419022, + 0.01933896914124489, + 0.015403283759951591, + 0.024135446175932884, + 0.0350986011326313, + 0.01658596098423004, + 0.02027226984500885, + 0.04164770245552063, + 0.028488919138908386, + 0.003862053854390979, + 0.002969121327623725, + 0.004703196696937084, + 0.004919483792036772, + 0.003060495713725686, + 0.0031130968127399683, + 0.003604406490921974, + 0.003405518364161253, + 0.0019471354316920042, + 0.001570486230775714, + 0.00629655085504055, + 0.005591697990894318, + 0.001949861878529191, + 0.0029594164807349443, + 0.0108945332467556, + 0.00678602047264576, + 0.0019128987332805991, + 0.0014443673426285386, + 0.0027632396668195724, + 0.0024499413557350636, + 0.001993716461583972, + 0.003567091189324856, + 0.0022515705786645412, + 0.002532458631321788, + 0.002362421015277505, + 0.003690826939418912, + 0.0017890171147882938, + 0.00279495632275939, + 0.0036728656850755215, + 0.0033805477432906628, + 0.002543854294344783, + 0.00195727008394897, + 0.009808885864913464, + 0.004553965292870998, + 0.013923206366598606, + 0.02361292764544487, + 0.012386837042868137, + 0.012680355459451675, + 0.14409130811691284, + 0.1343081146478653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.043053124099969864 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "single, static, high-angle, extreme-wide shot captures a catastrophic event, rendered entirely in Picasso's bold, fragmented, and multi-perspective style. A powerful mountain torrent surges across the landscape, not a flowing liquid but a turbulent collision of sharp, geometric planes in ochre browns and deep amber. This destructive force carves a path through the dense jungle of Java water cress, its delicate, parsley-like leaves deconstructed into jagged, acidic green shapes that churn and shatter within the Cubist vortex of roiling brown water. The scene is viewed simultaneously from multiple, impossible angles; we simultaneously see the flat surface of the floodwater, the cross-section of the water tearing through the soil, and the outlines of the waves. Against a dark, oppressive sky composed of scarred gray and black polygons, the raw power and destructive geometry of the event take center stage, creating a chaotic, flattened yet vibrant composition that emphasizes form and violent movement over realistic depiction.", + "seed": 11487, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_10/index_152_seed_11487_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09295056983828545, + "vlm_logprob": -0.09295056983828545, + "vlm_yes_no_margin": 2.699999952316284, + "vlm_token_credit": -0.09295056983828545, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1613200455904007 + ], + "opposite_logprob_tokens": [ + -1.9113200902938843 + ], + "logprob_reward": -0.1613200455904007, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1613200455904007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0951303094625473, + 0.018306612968444824, + 0.03519155830144882, + 0.5416688323020935, + 0.01749001443386078, + 0.03871624544262886, + 0.019385842606425285, + 0.02819802798330784, + 0.01099748071283102, + 0.1949150413274765 + ], + "credit_map": [ + 0.009999888949096203, + 0.01694403775036335, + 0.010158838704228401, + 0.010626301169395447, + 0.017259083688259125, + 0.008332718163728714, + 0.010782728902995586, + 0.011026705615222454, + 0.0013792632380500436, + 0.0020042143296450377, + 0.0028128644917160273, + 0.0035067901480942965, + 0.001644665258936584, + 0.0013657438103109598, + 0.002074889838695526, + 0.0035181811545044184, + 0.0021085285115987062, + 0.004895633086562157, + 0.005036624614149332, + 0.0056529901921749115, + 0.003275836119428277, + 0.0027730781584978104, + 0.007293079514056444, + 0.0041557857766747475, + 0.010628977790474892, + 0.01662028580904007, + 0.035171203315258026, + 0.1906757652759552, + 0.019659772515296936, + 0.009076875634491444, + 0.2371068149805069, + 0.022729141637682915, + 0.0012853493681177497, + 0.0012946047354489565, + 0.003424229333177209, + 0.003628087928518653, + 0.0012862500734627247, + 0.001386509626172483, + 0.0024608455132693052, + 0.0027241383213549852, + 0.0023488665465265512, + 0.0035611637867987156, + 0.011586177162826061, + 0.007648996077477932, + 0.0018875743262469769, + 0.003768601920455694, + 0.004433120600879192, + 0.003481739666312933, + 0.0015360494144260883, + 0.0018004908924922347, + 0.002459006616845727, + 0.004034063778817654, + 0.0016331482911482453, + 0.0010337155545130372, + 0.004593293182551861, + 0.002296075224876404, + 0.001915119239129126, + 0.002149993786588311, + 0.003944265190511942, + 0.003757108934223652, + 0.0014070643810555339, + 0.0019832258112728596, + 0.008217260241508484, + 0.004823989700525999, + 0.0010482504731044173, + 0.0010530027793720365, + 0.0016636515501886606, + 0.0023758942261338234, + 0.0007403685594908893, + 0.0011423222022131085, + 0.0015066320775076747, + 0.0014673590194433928, + 0.023105042055249214, + 0.014056136831641197, + 0.029498238116502762, + 0.029084650799632072, + 0.01742631196975708, + 0.01993943750858307, + 0.033054400235414505, + 0.028750816360116005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.1613200455904007 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358386501669884 + ], + "opposite_logprob_tokens": [ + -3.6513583660125732 + ], + "logprob_reward": -0.026358386501669884, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358386501669884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14193926751613617, + 0.026798319071531296, + 0.05411987751722336, + 0.5726668238639832, + 0.03527593985199928, + 0.0217383224517107, + 0.022905679419636726, + 0.020580677315592766, + 0.024764947593212128, + 0.07921012490987778 + ], + "credit_map": [ + 0.013748168013989925, + 0.022609680891036987, + 0.01837438903748989, + 0.016033882275223732, + 0.014600900001823902, + 0.014368916861712933, + 0.01871294341981411, + 0.023490384221076965, + 0.0021148896776139736, + 0.0029500899836421013, + 0.004160648211836815, + 0.005881589837372303, + 0.0027598990127444267, + 0.0024087384808808565, + 0.0028513125143945217, + 0.0036711518187075853, + 0.0035798242315649986, + 0.008112185634672642, + 0.007796379737555981, + 0.00907992385327816, + 0.00471154460683465, + 0.0038213569205254316, + 0.011051563546061516, + 0.005967100150883198, + 0.011526696383953094, + 0.017336683347821236, + 0.03782482445240021, + 0.20012615621089935, + 0.02147400937974453, + 0.010159125551581383, + 0.25100889801979065, + 0.02321038953959942, + 0.0024385915603488684, + 0.002568097086623311, + 0.007946653291583061, + 0.006920486688613892, + 0.0028619347140192986, + 0.003630084916949272, + 0.004223036114126444, + 0.004687055014073849, + 0.0011052123736590147, + 0.002171838888898492, + 0.0033808029256761074, + 0.0035213958472013474, + 0.004401097074151039, + 0.0019168427679687738, + 0.0026727693621069193, + 0.002568363444879651, + 0.0016076324973255396, + 0.00208488036878407, + 0.0032532347831875086, + 0.0035203348379582167, + 0.0023091568145900965, + 0.002166441874578595, + 0.004400425124913454, + 0.003563572419807315, + 0.0015159028116613626, + 0.0025682239793241024, + 0.003877076553180814, + 0.0035429191775619984, + 0.0013626428553834558, + 0.00157490826677531, + 0.0031202530954033136, + 0.003018750576302409, + 0.001543142250739038, + 0.0018780624959617853, + 0.003894168883562088, + 0.005536589305847883, + 0.0019019792089238763, + 0.0028718344401568174, + 0.004007528070360422, + 0.0031316429376602173, + 0.008192094042897224, + 0.006103802938014269, + 0.013065124861896038, + 0.013588020578026772, + 0.006777157075703144, + 0.008246739394962788, + 0.012158770114183426, + 0.011078409850597382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.026358386501669884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181953813880682 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.0181953813880682, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181953813880682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16281607747077942, + 0.13527636229991913, + 0.10167227685451508, + 0.30001792311668396, + 0.06132932007312775, + 0.023825300857424736, + 0.03241310641169548, + 0.03909151628613472, + 0.023864012211561203, + 0.11969415098428726 + ], + "credit_map": [ + 0.01825970783829689, + 0.02236403524875641, + 0.015909722074866295, + 0.017520884051918983, + 0.03439105302095413, + 0.013565653003752232, + 0.020322205498814583, + 0.020482799038290977, + 0.010064919479191303, + 0.013315313495695591, + 0.020944328978657722, + 0.03631419315934181, + 0.009561246261000633, + 0.0075964489951729774, + 0.022640079259872437, + 0.014839829877018929, + 0.009226332418620586, + 0.014813626185059547, + 0.014644580893218517, + 0.013194996863603592, + 0.01065122801810503, + 0.010214559733867645, + 0.01861877366900444, + 0.010308167897164822, + 0.009802328422665596, + 0.012033660896122456, + 0.02206379920244217, + 0.09107030183076859, + 0.019067667424678802, + 0.011100465431809425, + 0.11911489814519882, + 0.015764795243740082, + 0.0068866414949297905, + 0.005749407224357128, + 0.010327217169106007, + 0.010637680999934673, + 0.004729248117655516, + 0.006845192052423954, + 0.007271890994161367, + 0.008882030844688416, + 0.0015780602116137743, + 0.0022356382105499506, + 0.004697788041085005, + 0.004423141945153475, + 0.0028293856885284185, + 0.002313982928171754, + 0.0027706443797796965, + 0.002976658521220088, + 0.002880326472222805, + 0.003338347654789686, + 0.0037550029810518026, + 0.006070542614907026, + 0.004441410303115845, + 0.0028519625775516033, + 0.005345527082681656, + 0.0037299839314073324, + 0.0028041861951351166, + 0.0032051862217485905, + 0.005496974103152752, + 0.007256136275827885, + 0.003238934325054288, + 0.0036765537224709988, + 0.008485707454383373, + 0.004927833564579487, + 0.002683142200112343, + 0.003952182363718748, + 0.003022122662514448, + 0.0037259459495544434, + 0.002808448625728488, + 0.0022816394921392202, + 0.0024534803815186024, + 0.0029370479751378298, + 0.013983809389173985, + 0.009935611858963966, + 0.01668117195367813, + 0.018483567982912064, + 0.011959443800151348, + 0.013486539013683796, + 0.017102301120758057, + 0.018061690032482147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0181953813880682 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991624772548676 + ], + "opposite_logprob_tokens": [ + -1.804916262626648 + ], + "logprob_reward": -0.17991624772548676, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991624772548676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12244460731744766, + 0.0782940536737442, + 0.0993732288479805, + 0.5305535793304443, + 0.05614253878593445, + 0.0165412537753582, + 0.017121952027082443, + 0.015934990718960762, + 0.016764013096690178, + 0.04682979732751846 + ], + "credit_map": [ + 0.016358526423573494, + 0.01368777547031641, + 0.012415681965649128, + 0.011605886742472649, + 0.022859923541545868, + 0.010514257475733757, + 0.01770932413637638, + 0.017293229699134827, + 0.0061349691823124886, + 0.006508910562843084, + 0.00981887150555849, + 0.015067419968545437, + 0.00681711919605732, + 0.006079074926674366, + 0.019229160621762276, + 0.008638516068458557, + 0.009174314327538013, + 0.01566285640001297, + 0.01446455717086792, + 0.01464130636304617, + 0.009528367780148983, + 0.008011208847165108, + 0.018335914239287376, + 0.009554688818752766, + 0.015278460457921028, + 0.01929899863898754, + 0.03373492509126663, + 0.17487585544586182, + 0.028186576440930367, + 0.0161758903414011, + 0.21857059001922607, + 0.024432264268398285, + 0.0039589786902070045, + 0.0032967824954539537, + 0.009101437404751778, + 0.0130019411444664, + 0.004970635287463665, + 0.006278655491769314, + 0.006793894339352846, + 0.008740205317735672, + 0.0009815499652177095, + 0.0022076782770454884, + 0.002948167035356164, + 0.0028923682402819395, + 0.003434394486248493, + 0.001248978078365326, + 0.0014127646572887897, + 0.0014153512893244624, + 0.0013558784266933799, + 0.0015583689091727138, + 0.002141042146831751, + 0.0033352079335600138, + 0.0021624132059514523, + 0.001659088535234332, + 0.0028054669965058565, + 0.002104485174641013, + 0.0012106671929359436, + 0.001757525373250246, + 0.0029946023132652044, + 0.0020295330323278904, + 0.0012307745637372136, + 0.00141697668004781, + 0.0024568450171500444, + 0.0028380637522786856, + 0.0023029684089124203, + 0.002839040709659457, + 0.0020243965554982424, + 0.001813572715036571, + 0.0025367755442857742, + 0.0019968156702816486, + 0.0017522310372442007, + 0.0014982112916186452, + 0.004320702515542507, + 0.004484407138079405, + 0.007854868657886982, + 0.005681856069713831, + 0.006436431780457497, + 0.0052840872667729855, + 0.006902584806084633, + 0.005864856299012899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.17991624772548676 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a Picasso-style Cubist aesthetic with bold, fragmented, and multi-perspective geometry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject a powerful mountain torrent depicted as a collision of sharp geometric planes in ochre browns and deep amber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Java water cress plants shown as delicate, parsley-like leaves deconstructed into jagged, acidic green shapes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and oppressive, composed of scarred gray and black polygons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a single, static, high-angle, extreme-wide view showing the scene from multiple simultaneous impossible angles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789627879858017 + ], + "opposite_logprob_tokens": [ + -2.578962802886963 + ], + "logprob_reward": -0.0789627879858017, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789627879858017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20268625020980835, + 0.10253375768661499, + 0.08941145986318588, + 0.08691862225532532, + 0.17321033775806427, + 0.10011132806539536, + 0.027457907795906067, + 0.0261274091899395, + 0.0167706161737442, + 0.17477239668369293 + ], + "credit_map": [ + 0.023790182545781136, + 0.027461104094982147, + 0.021873963996767998, + 0.030451349914073944, + 0.024865834042429924, + 0.026817260310053825, + 0.022359542548656464, + 0.025067009031772614, + 0.00812296848744154, + 0.008357522077858448, + 0.017194781452417374, + 0.023558268323540688, + 0.01057898998260498, + 0.00818814616650343, + 0.013775988481938839, + 0.012757088989019394, + 0.006407732143998146, + 0.01015617698431015, + 0.013007784262299538, + 0.013129735365509987, + 0.008130913600325584, + 0.007261188235133886, + 0.01958342082798481, + 0.011734512634575367, + 0.0030918747652322054, + 0.0034374771639704704, + 0.007760107051581144, + 0.025563515722751617, + 0.004544417839497328, + 0.003918481059372425, + 0.03277822583913803, + 0.005824524909257889, + 0.01189387682825327, + 0.0119919553399086, + 0.030042899772524834, + 0.037040822207927704, + 0.01727185770869255, + 0.019341962411999702, + 0.02113937959074974, + 0.024487577378749847, + 0.006050007417798042, + 0.013936481438577175, + 0.022415099665522575, + 0.017803223803639412, + 0.01188752893358469, + 0.010370524600148201, + 0.009822116233408451, + 0.007826345972716808, + 0.0021396714728325605, + 0.003283029654994607, + 0.0033104054164141417, + 0.0064665405079722404, + 0.003310298314318061, + 0.0018075787229463458, + 0.003646073630079627, + 0.00349431112408638, + 0.0023144378792494535, + 0.0017324908403679729, + 0.003947488963603973, + 0.0036477616522461176, + 0.0021112116519361734, + 0.0019332742085680366, + 0.006386262830346823, + 0.004054481163620949, + 0.000997176393866539, + 0.0027484940364956856, + 0.0018853769870474935, + 0.0016442592022940516, + 0.003148067509755492, + 0.002301495522260666, + 0.002019674051553011, + 0.002026073634624481, + 0.0187455415725708, + 0.011973114684224129, + 0.026923535391688347, + 0.02359776198863983, + 0.015422931872308254, + 0.017046388238668442, + 0.03764110058546066, + 0.02342202328145504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0789627879858017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_11/index_tensor([5422, 5422])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_11/index_tensor([5422, 5422])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f39ca255a5252e7512bb84c41cfba3ffea43bc01 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_11/index_tensor([5422, 5422])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "It's an extreme close-up, shot with a still camera in realistic, high-speed slow motion, focusing on a lightweight paper receipt resting on a dark, polished granite kitchen countertop. Made of thin, off-white thermal paper, the receipt is slightly wrinkled, with sharp creases and lines of text printed in faded, dark ink visible but barely discernible. A gentle, unseen breeze, originating off-camera, begins cautiously lifting one edge of the paper. The receipt rises in exquisite slow motion, undulating in an elegant, almost liquid dance, its thin body twisting and fluttering inches above the smooth surface, revealing every subtle ripple and twist as it momentarily hovers in mid-air. As the air currents subside, the receipt begins to descend, deliberately floating down, landing back on the countertop, not flat but with a soft, final curl, one corner touching first, before the rest of the paper gently collapses. Soft morning light streams across the scene from a nearby window, capturing the fluttering edges of the paper with bright highlights and casting long, soft, dynamic shadows that move with every movement of the receipt, creating a serene, contemplative, and visually mesmerizing atmosphere, completely focused on the delicate physics of this ordinary moment.", + "seed": 22563, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_11/index_5422_seed_22563_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13187946313992144, + "vlm_logprob": -0.13187946313992144, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.13187946313992144, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043448928743600845 + ], + "opposite_logprob_tokens": [ + -3.4184489250183105 + ], + "logprob_reward": -0.043448928743600845, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.043448928743600845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13235875964164734, + 0.13788102567195892, + 0.07887711375951767, + 0.05534343421459198, + 0.09932955354452133, + 0.01441353652626276, + 0.020972969010472298, + 0.320365309715271, + 0.04943317174911499, + 0.09102510660886765 + ], + "credit_map": [ + 0.012748237699270248, + 0.01949313096702099, + 0.010864107869565487, + 0.013551658019423485, + 0.01743241399526596, + 0.009919191710650921, + 0.020010756328701973, + 0.028339266777038574, + 0.019594775512814522, + 0.01081469003111124, + 0.011656027287244797, + 0.014809180982410908, + 0.020288540050387383, + 0.012372776865959167, + 0.02090568281710148, + 0.02743934467434883, + 0.012034676037728786, + 0.007404385134577751, + 0.008929155766963959, + 0.008117198944091797, + 0.008347229100763798, + 0.014652159996330738, + 0.01145448163151741, + 0.007937823422253132, + 0.006560203619301319, + 0.006333861034363508, + 0.006688263732939959, + 0.007543676532804966, + 0.007241240702569485, + 0.00877466145902872, + 0.006722831167280674, + 0.005478698760271072, + 0.01408211886882782, + 0.010875634849071503, + 0.013325389474630356, + 0.01514173112809658, + 0.0061849975027143955, + 0.015474244020879269, + 0.014647559262812138, + 0.009597880765795708, + 0.0020591162610799074, + 0.001455864286981523, + 0.0016690210904926062, + 0.0020625190809369087, + 0.0012528834631666541, + 0.0020777718164026737, + 0.002037776168435812, + 0.0017985844751819968, + 0.0028129664715379477, + 0.0030281497165560722, + 0.0033767109271138906, + 0.002349550137296319, + 0.002004319103434682, + 0.004063666798174381, + 0.001351631828583777, + 0.00198597414419055, + 0.0417916439473629, + 0.03952982649207115, + 0.010062542743980885, + 0.008874342776834965, + 0.10313429683446884, + 0.07194370776414871, + 0.017740460112690926, + 0.027288522571325302, + 0.007185245398432016, + 0.005915943067520857, + 0.005760572385042906, + 0.004005569498986006, + 0.008057492785155773, + 0.010527546517550945, + 0.0029191828798502684, + 0.005061618518084288, + 0.01280964445322752, + 0.009886661544442177, + 0.012205752544105053, + 0.011357908137142658, + 0.010822458192706108, + 0.011158290319144726, + 0.013671613298356533, + 0.00911277811974287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.043448928743600845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358386501669884 + ], + "opposite_logprob_tokens": [ + -3.6513583660125732 + ], + "logprob_reward": -0.026358386501669884, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358386501669884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16799810528755188, + 0.07834906131029129, + 0.058670103549957275, + 0.06823106110095978, + 0.057912059128284454, + 0.04275060445070267, + 0.16046030819416046, + 0.2565877139568329, + 0.04384802281856537, + 0.06519293040037155 + ], + "credit_map": [ + 0.01966581679880619, + 0.025766834616661072, + 0.014431806281208992, + 0.012559141032397747, + 0.013352442532777786, + 0.015112061984837055, + 0.029294725507497787, + 0.03781528025865555, + 0.009095950052142143, + 0.010758849792182446, + 0.00816414412111044, + 0.008566703647375107, + 0.009388480335474014, + 0.00693265488371253, + 0.011180753819644451, + 0.014261522330343723, + 0.00957091711461544, + 0.006056802812963724, + 0.005475577898323536, + 0.00452036177739501, + 0.007016844116151333, + 0.013786381110548973, + 0.006626768968999386, + 0.005616450682282448, + 0.010795487090945244, + 0.00797563511878252, + 0.011172410100698471, + 0.007599841337651014, + 0.0059258826076984406, + 0.009874120354652405, + 0.007333043031394482, + 0.007554641924798489, + 0.008284248411655426, + 0.0069619640707969666, + 0.00838249921798706, + 0.006163835059851408, + 0.004742303863167763, + 0.011470004916191101, + 0.0063879359513521194, + 0.005519272293895483, + 0.005675052758306265, + 0.003847244195640087, + 0.0044919527135789394, + 0.007524640299379826, + 0.004185519181191921, + 0.007262181956321001, + 0.004733963869512081, + 0.005030050408095121, + 0.021114131435751915, + 0.02436680532991886, + 0.015843212604522705, + 0.007843957282602787, + 0.027721459046006203, + 0.04440649226307869, + 0.00737253250554204, + 0.011791721917688847, + 0.033465273678302765, + 0.02699519507586956, + 0.010659468360245228, + 0.011678002774715424, + 0.08064430207014084, + 0.05990903824567795, + 0.013571820221841335, + 0.019664598628878593, + 0.005853178910911083, + 0.006722502410411835, + 0.004229637328535318, + 0.0057018171064555645, + 0.006544176023453474, + 0.006127211265265942, + 0.004724397324025631, + 0.0039451029151678085, + 0.006827878765761852, + 0.007343775127083063, + 0.007106533739715815, + 0.010706514120101929, + 0.009044762700796127, + 0.011169771663844585, + 0.006962307263165712, + 0.006031387485563755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.026358386501669884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761901140213013 + ], + "opposite_logprob_tokens": [ + -0.8261901140213013 + ], + "logprob_reward": -0.5761901140213013, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761901140213013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07596034556627274, + 0.05046750232577324, + 0.06134852394461632, + 0.08868467062711716, + 0.15220977365970612, + 0.0853055939078331, + 0.03590409830212593, + 0.3202635943889618, + 0.10955101251602173, + 0.02030489221215248 + ], + "credit_map": [ + 0.008858836255967617, + 0.011741545051336288, + 0.011593443341553211, + 0.005743693094700575, + 0.009441526606678963, + 0.006718803197145462, + 0.01248160656541586, + 0.009380892850458622, + 0.005369937047362328, + 0.004754634574055672, + 0.002749964827671647, + 0.004553086590021849, + 0.006170390639454126, + 0.004468953236937523, + 0.007504754234105349, + 0.014895781874656677, + 0.008462906815111637, + 0.004983508959412575, + 0.0073981680907309055, + 0.006298992782831192, + 0.006028290372341871, + 0.012871822342276573, + 0.009408537298440933, + 0.005896296817809343, + 0.013037508353590965, + 0.006583759561181068, + 0.019207431003451347, + 0.009757766500115395, + 0.005769574549049139, + 0.013422330841422081, + 0.010622985661029816, + 0.010283319279551506, + 0.024077897891402245, + 0.016154933720827103, + 0.02913321927189827, + 0.019068831577897072, + 0.012492317706346512, + 0.017334679141640663, + 0.01788409799337387, + 0.016063790768384933, + 0.013656584545969963, + 0.01086356956511736, + 0.01159091480076313, + 0.014679384417831898, + 0.007897976785898209, + 0.01108288113027811, + 0.008897536434233189, + 0.006636742036789656, + 0.0048059821128845215, + 0.0046971626579761505, + 0.003153002355247736, + 0.0024032900109887123, + 0.005330721847712994, + 0.009179649874567986, + 0.0030110483057796955, + 0.0033232395071536303, + 0.04292050004005432, + 0.0372718870639801, + 0.011359657160937786, + 0.010156224481761456, + 0.10544393956661224, + 0.07215684652328491, + 0.0183312576264143, + 0.022623291239142418, + 0.01195381861180067, + 0.013740712776780128, + 0.009371697902679443, + 0.014700408093631268, + 0.019653018563985825, + 0.015986479818820953, + 0.010618658736348152, + 0.01352621428668499, + 0.0032323356717824936, + 0.0036280741915106773, + 0.002630258444696665, + 0.002149771200492978, + 0.0023813252337276936, + 0.002505342476069927, + 0.0020857660565525293, + 0.001692018355242908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.5761901140213013 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003618124406784773 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.003618124406784773, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003618124406784773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1290656328201294, + 0.1245613545179367, + 0.060573283582925797, + 0.04238134250044823, + 0.03711266070604324, + 0.025157691910862923, + 0.08096282184123993, + 0.3184840679168701, + 0.040822964161634445, + 0.1408781260251999 + ], + "credit_map": [ + 0.014439299702644348, + 0.0175261739641428, + 0.016036072745919228, + 0.011023280210793018, + 0.016464872285723686, + 0.0104300482198596, + 0.025691363960504532, + 0.01745452918112278, + 0.01824377104640007, + 0.015946848317980766, + 0.01069606002420187, + 0.011418649926781654, + 0.015377226285636425, + 0.012498107738792896, + 0.015587979927659035, + 0.02479272149503231, + 0.010121747851371765, + 0.00584749411791563, + 0.007926734164357185, + 0.0062476592138409615, + 0.0058092693798244, + 0.009368650615215302, + 0.009063834324479103, + 0.006187892984598875, + 0.0059835887514054775, + 0.0033485949970781803, + 0.004722675774246454, + 0.004164102021604776, + 0.004200930241495371, + 0.009914660826325417, + 0.006210335996001959, + 0.003836456686258316, + 0.005407517775893211, + 0.003327046986669302, + 0.006731266621500254, + 0.005283381324261427, + 0.003738570725545287, + 0.004523454234004021, + 0.004744237754493952, + 0.00335718528367579, + 0.0029002721421420574, + 0.003101060166954994, + 0.0029214743990451097, + 0.003931096754968166, + 0.002873528515920043, + 0.004489582497626543, + 0.0028117727488279343, + 0.0021289032883942127, + 0.009456777945160866, + 0.008829496800899506, + 0.008948556147515774, + 0.008955935947597027, + 0.010022026486694813, + 0.015795595943927765, + 0.01062393095344305, + 0.008330504409968853, + 0.04087090492248535, + 0.03859688341617584, + 0.011642411351203918, + 0.010414495132863522, + 0.10378212481737137, + 0.07203053683042526, + 0.018360799178481102, + 0.022785881534218788, + 0.006044204346835613, + 0.0035624695010483265, + 0.0038678732234984636, + 0.0049027069471776485, + 0.005902663804590702, + 0.0076703294180333614, + 0.0037555363960564137, + 0.005117184016853571, + 0.015976862981915474, + 0.018764449283480644, + 0.010712901130318642, + 0.014810612425208092, + 0.02780700847506523, + 0.030226537957787514, + 0.011088425293564796, + 0.011491334065794945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.003618124406784773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781762026250362 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781762026250362, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781762026250362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15025007724761963, + 0.1966591477394104, + 0.03869681805372238, + 0.0497705452144146, + 0.02468433603644371, + 0.029258813709020615, + 0.11488481611013412, + 0.3454856872558594, + 0.03343292698264122, + 0.016876813024282455 + ], + "credit_map": [ + 0.016302440315485, + 0.02479841187596321, + 0.0192502923309803, + 0.01501645240932703, + 0.02019651047885418, + 0.012906484305858612, + 0.020689725875854492, + 0.021089769899845123, + 0.021742157638072968, + 0.032593026757240295, + 0.030843472108244896, + 0.022417288273572922, + 0.024202777072787285, + 0.02111797221004963, + 0.020201006904244423, + 0.023541435599327087, + 0.006864180788397789, + 0.0041874852031469345, + 0.004577104467898607, + 0.0033141723833978176, + 0.004277867265045643, + 0.007383941672742367, + 0.004277603235095739, + 0.003814466530457139, + 0.006322260946035385, + 0.006158029194921255, + 0.006423390470445156, + 0.0039124079048633575, + 0.005262268707156181, + 0.013184154406189919, + 0.004510931670665741, + 0.003997100982815027, + 0.004092451650649309, + 0.0021872713696211576, + 0.007144037168473005, + 0.002253604354336858, + 0.0016647391021251678, + 0.0028021344915032387, + 0.0020377750042825937, + 0.0025023226626217365, + 0.004574696067720652, + 0.0032672835513949394, + 0.004917412530630827, + 0.004662788938730955, + 0.0025003082118928432, + 0.004099591169506311, + 0.002780389040708542, + 0.0024563446640968323, + 0.01791822351515293, + 0.01674540899693966, + 0.019151022657752037, + 0.006996504962444305, + 0.01411738246679306, + 0.021521029993891716, + 0.00935363955795765, + 0.009081598371267319, + 0.04351110756397247, + 0.03337721899151802, + 0.011371668428182602, + 0.010488775558769703, + 0.11448914557695389, + 0.08455998450517654, + 0.019028862938284874, + 0.02865893580019474, + 0.004421463701874018, + 0.005049047991633415, + 0.002784280339255929, + 0.0033072649966925383, + 0.005709492601454258, + 0.006579931825399399, + 0.002207196783274412, + 0.003374251304194331, + 0.0023229161743074656, + 0.0022667022421956062, + 0.0017437811475247145, + 0.002263423288241029, + 0.002366337226703763, + 0.0025168368592858315, + 0.0018775061471387744, + 0.0015193111030384898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.009781762026250362 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "It's an extreme close-up, shot with a still camera in realistic, high-speed slow motion, focusing on a lightweight paper receipt resting on a dark, polished granite kitchen countertop. Made of thin, off-white thermal paper, the receipt is slightly wrinkled, with sharp creases and lines of text printed in faded, dark ink visible but barely discernible. A gentle, unseen breeze, originating off-camera, begins cautiously lifting one edge of the paper. The receipt rises in exquisite slow motion, undulating in an elegant, almost liquid dance, its thin body twisting and fluttering inches above the smooth surface, revealing every subtle ripple and twist as it momentarily hovers in mid-air. As the air currents subside, the receipt begins to descend, deliberately floating down, landing back on the countertop, not flat but with a soft, final curl, one corner touching first, before the rest of the paper gently collapses. Soft morning light streams across the scene from a nearby window, capturing the fluttering edges of the paper with bright highlights and casting long, soft, dynamic shadows that move with every movement of the receipt, creating a serene, contemplative, and visually mesmerizing atmosphere, completely focused on the delicate physics of this ordinary moment.", + "seed": 22564, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_11/index_5422_seed_22564_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.24668162474408745, + "vlm_logprob": -0.24668162474408745, + "vlm_yes_no_margin": 2.775000011920929, + "vlm_token_credit": -0.24668162474408745, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12234273552894592 + ], + "opposite_logprob_tokens": [ + -2.247342824935913 + ], + "logprob_reward": -0.12234273552894592, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12234273552894592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046578098088502884, + 0.4851881265640259, + 0.0627555400133133, + 0.07055488973855972, + 0.19057798385620117, + 0.028294706717133522, + 0.019556192681193352, + 0.023548509925603867, + 0.04701325297355652, + 0.025932786986231804 + ], + "credit_map": [ + 0.0028693336062133312, + 0.005006279330700636, + 0.0035585258156061172, + 0.007768855430185795, + 0.006841578986495733, + 0.005200852174311876, + 0.007203711196780205, + 0.008128957822918892, + 0.030352240428328514, + 0.03323953226208687, + 0.10357508808374405, + 0.02694755233824253, + 0.01987573131918907, + 0.17230090498924255, + 0.05329246446490288, + 0.04560455307364464, + 0.003983647096902132, + 0.005227045156061649, + 0.010183693841099739, + 0.0046396092511713505, + 0.0066003622487187386, + 0.01585410349071026, + 0.007707221433520317, + 0.00855985190719366, + 0.004087844397872686, + 0.01337909884750843, + 0.007275549694895744, + 0.008201672695577145, + 0.00634482316672802, + 0.016657138243317604, + 0.007223909255117178, + 0.0073848566971719265, + 0.010376319289207458, + 0.06119531765580177, + 0.03067374788224697, + 0.019329844042658806, + 0.014559689909219742, + 0.019644318148493767, + 0.01829519122838974, + 0.01650354266166687, + 0.002568810945376754, + 0.004431094508618116, + 0.0038439466152340174, + 0.005594449117779732, + 0.0019172955071553588, + 0.003670069621875882, + 0.0030592528637498617, + 0.0032097853254526854, + 0.0013542396482080221, + 0.005606509279459715, + 0.001931992475874722, + 0.0015676863258704543, + 0.005010006949305534, + 0.0014169295318424702, + 0.0012639339547604322, + 0.0014048925368115306, + 0.0014011916937306523, + 0.003972124308347702, + 0.0016744858585298061, + 0.0027426383458077908, + 0.005909200292080641, + 0.00182019779458642, + 0.0025509835686534643, + 0.00347768678329885, + 0.003339440794661641, + 0.006070232950150967, + 0.004038802348077297, + 0.007175246253609657, + 0.007337682880461216, + 0.003869669744744897, + 0.007321668788790703, + 0.007860508747398853, + 0.0017022003885358572, + 0.004782089963555336, + 0.0020831204019486904, + 0.003156098537147045, + 0.005063031800091267, + 0.002606776775792241, + 0.003616489004343748, + 0.0029229780193418264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.12234273552894592 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029809314757585526 + ], + "opposite_logprob_tokens": [ + -3.5298092365264893 + ], + "logprob_reward": -0.029809314757585526, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809314757585526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1861659288406372, + 0.14483393728733063, + 0.021856101229786873, + 0.0540715754032135, + 0.1392158567905426, + 0.028392435982823372, + 0.01603735238313675, + 0.0265831146389246, + 0.0721294954419136, + 0.3107142448425293 + ], + "credit_map": [ + 0.01006112340837717, + 0.01438582967966795, + 0.014107716269791126, + 0.02413298934698105, + 0.026222942396998405, + 0.023833902552723885, + 0.037573061883449554, + 0.03584834933280945, + 0.009059607982635498, + 0.009625178761780262, + 0.030370725318789482, + 0.008929566480219364, + 0.006724669132381678, + 0.05082985758781433, + 0.015687452629208565, + 0.013606879860162735, + 0.0022561002988368273, + 0.001984265400096774, + 0.004806140903383493, + 0.001782144419848919, + 0.0017910065362229943, + 0.004385699983686209, + 0.002335586817935109, + 0.0025151569861918688, + 0.003079813439399004, + 0.021139811724424362, + 0.0096632344648242, + 0.0031643591355532408, + 0.002922971500083804, + 0.0074624777771532536, + 0.0024288499262183905, + 0.004210061859339476, + 0.005670590326189995, + 0.05903184786438942, + 0.02689414843916893, + 0.007172985002398491, + 0.007638380862772465, + 0.020515356212854385, + 0.0048933387733995914, + 0.007399219088256359, + 0.002355503384023905, + 0.005454725585877895, + 0.003075723070651293, + 0.005518406629562378, + 0.0022734608501195908, + 0.003751303069293499, + 0.003007760038599372, + 0.0029555526562035084, + 0.0009553100098855793, + 0.003717839252203703, + 0.0012579713948071003, + 0.0015081154415383935, + 0.003538202727213502, + 0.0016834799898788333, + 0.001386051531881094, + 0.0019903816282749176, + 0.0027294771280139685, + 0.0035557083319872618, + 0.0030037094838917255, + 0.00390980439260602, + 0.0034671081230044365, + 0.002975664334371686, + 0.002837009495124221, + 0.004104632884263992, + 0.005672826897352934, + 0.011073291301727295, + 0.00721606845036149, + 0.014540358446538448, + 0.009258942678570747, + 0.007154188118875027, + 0.007812508381903172, + 0.009401308372616768, + 0.022984638810157776, + 0.07856600731611252, + 0.025700774043798447, + 0.04622020199894905, + 0.0561540313065052, + 0.03470868244767189, + 0.024171587079763412, + 0.02220829203724861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.029809314757585526 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539216995239258 + ], + "opposite_logprob_tokens": [ + -0.42892175912857056 + ], + "logprob_reward": -1.0539216995239258, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0539216995239258, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0477818064391613, + 0.03608463332056999, + 0.018916543573141098, + 0.5361262559890747, + 0.3120253384113312, + 0.010987106710672379, + 0.012855175882577896, + 0.0036566148046404123, + 0.012321357615292072, + 0.009245202876627445 + ], + "credit_map": [ + 0.002979427110403776, + 0.0037706242874264717, + 0.007393152453005314, + 0.005360310431569815, + 0.006150895729660988, + 0.01201087050139904, + 0.005372770596295595, + 0.004743759986013174, + 0.002303031273186207, + 0.0029877631459385157, + 0.006621226668357849, + 0.002738625044003129, + 0.00250108540058136, + 0.011354335583746433, + 0.004426561761647463, + 0.0031520065385848284, + 0.0015438037225976586, + 0.002012574812397361, + 0.0021731879096478224, + 0.0018716261256486177, + 0.0031310622580349445, + 0.00233337446115911, + 0.0032455921173095703, + 0.0026053222827613354, + 0.02173428237438202, + 0.2299983650445938, + 0.09351354837417603, + 0.022711895406246185, + 0.029704885557293892, + 0.09218697994947433, + 0.013673008419573307, + 0.03260323777794838, + 0.013032487593591213, + 0.13075904548168182, + 0.05753171071410179, + 0.016808951273560524, + 0.018312012776732445, + 0.04829033464193344, + 0.01152210682630539, + 0.01576869562268257, + 0.0009745038696564734, + 0.002883337205275893, + 0.0014689028030261397, + 0.0008998974808491766, + 0.0008723103674128652, + 0.00135536037851125, + 0.0010975938057526946, + 0.001435201964341104, + 0.0006091752438805997, + 0.005202538333833218, + 0.001488382345996797, + 0.0006882076268084347, + 0.0028739278204739094, + 0.0009056531707756221, + 0.0005460717948153615, + 0.0005412203026935458, + 0.0002796580665744841, + 0.0005258990568108857, + 0.0003159720217809081, + 0.0008141365251503885, + 0.00037500588223338127, + 0.00039897390524856746, + 0.0004500614886637777, + 0.0004969078581780195, + 0.0009509935625828803, + 0.0019233094062656164, + 0.0009778114035725594, + 0.0016044541262090206, + 0.0020561895798891783, + 0.0009318781667388976, + 0.002004689769819379, + 0.001872032298706472, + 0.0007190160686150193, + 0.0016143769025802612, + 0.0009832235518842936, + 0.0010407381923869252, + 0.0016987137496471405, + 0.0009196040919050574, + 0.001255336683243513, + 0.0010141939856112003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -1.0539216995239258 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744359154254198 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.006744359154254198, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744359154254198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10945556312799454, + 0.11215230077505112, + 0.06099816411733627, + 0.39287811517715454, + 0.10924675315618515, + 0.12255626171827316, + 0.016895722597837448, + 0.020976250991225243, + 0.028642818331718445, + 0.026198051869869232 + ], + "credit_map": [ + 0.009007558226585388, + 0.01775439828634262, + 0.012418230064213276, + 0.01193676795810461, + 0.017830846831202507, + 0.016246650367975235, + 0.01367969810962677, + 0.010581409558653831, + 0.006902812048792839, + 0.0073349568992853165, + 0.022830715402960777, + 0.008425047621130943, + 0.006307157687842846, + 0.03679049015045166, + 0.012581026181578636, + 0.010980096645653248, + 0.003570237895473838, + 0.005740061402320862, + 0.008942593820393085, + 0.0041983784176409245, + 0.005293886177241802, + 0.017207862809300423, + 0.007699556183069944, + 0.00834558717906475, + 0.01669527404010296, + 0.1671208292245865, + 0.07351327687501907, + 0.017628885805606842, + 0.02222469076514244, + 0.060259636491537094, + 0.010961242951452732, + 0.024474291130900383, + 0.004557769745588303, + 0.046075787395238876, + 0.020657319575548172, + 0.005564277991652489, + 0.006172848865389824, + 0.01697133108973503, + 0.0035516235511749983, + 0.005695797502994537, + 0.008717712946236134, + 0.02834540233016014, + 0.01714615523815155, + 0.020214039832353592, + 0.009700184687972069, + 0.015775976702570915, + 0.010604697279632092, + 0.012052091769874096, + 0.0010270572965964675, + 0.0042664045467972755, + 0.001395322964526713, + 0.001649065176025033, + 0.0037509251851588488, + 0.001191100338473916, + 0.0019157836213707924, + 0.0017000642837956548, + 0.0013065218226984143, + 0.002896843943744898, + 0.0014553035143762827, + 0.0033441493287682533, + 0.004031028598546982, + 0.0019205036805942655, + 0.0023159354459494352, + 0.003705964656546712, + 0.001642051269300282, + 0.0036428372841328382, + 0.002144980477169156, + 0.0035536037757992744, + 0.007428115699440241, + 0.0040262495167553425, + 0.002760589122772217, + 0.0034443901386111975, + 0.0018844626611098647, + 0.006539303343743086, + 0.002492809435352683, + 0.0038132364861667156, + 0.004528311546891928, + 0.002845499897375703, + 0.0025237188674509525, + 0.0015707073034718633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.006744359154254198 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059001475572586 + ], + "opposite_logprob_tokens": [ + -3.89559006690979 + ], + "logprob_reward": -0.02059001475572586, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059001475572586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12041568756103516, + 0.11430833488702774, + 0.04981450363993645, + 0.09163298457860947, + 0.43225201964378357, + 0.08686666935682297, + 0.02222520485520363, + 0.02350325882434845, + 0.01942444033920765, + 0.0395568385720253 + ], + "credit_map": [ + 0.009150252677500248, + 0.00764483492821455, + 0.01423689629882574, + 0.017011437565088272, + 0.014397396706044674, + 0.01955525577068329, + 0.01933172531425953, + 0.0190879013389349, + 0.0066980887204408646, + 0.008879246190190315, + 0.020318586379289627, + 0.009445934556424618, + 0.007640584837645292, + 0.03599344566464424, + 0.012960250489413738, + 0.01237220037728548, + 0.0036381909158080816, + 0.0043436759151518345, + 0.009192348457872868, + 0.0037833910901099443, + 0.0044573466293513775, + 0.012785605154931545, + 0.005652994848787785, + 0.005960952956229448, + 0.005079476162791252, + 0.031279176473617554, + 0.005516214296221733, + 0.0037454329431056976, + 0.004740015137940645, + 0.02643676847219467, + 0.004865752998739481, + 0.00997015368193388, + 0.017783472314476967, + 0.18712158501148224, + 0.08638183027505875, + 0.018719051033258438, + 0.021868852898478508, + 0.0666840672492981, + 0.011241969652473927, + 0.02245120331645012, + 0.010321259498596191, + 0.012261620722711086, + 0.011240297928452492, + 0.011623634956777096, + 0.006773605477064848, + 0.012654241174459457, + 0.009407893754541874, + 0.012584124691784382, + 0.0025092342402786016, + 0.003363901749253273, + 0.003326758509501815, + 0.002324024448171258, + 0.006023415829986334, + 0.001751313335262239, + 0.001295257592573762, + 0.0016313001979142427, + 0.001525679137557745, + 0.0032808459363877773, + 0.001948427059687674, + 0.003832679009065032, + 0.005441905930638313, + 0.0019480373011901975, + 0.00251684058457613, + 0.00300884828902781, + 0.0015943950274959207, + 0.0028536885511130095, + 0.0018266828265041113, + 0.00374895753338933, + 0.003014462301507592, + 0.002173943677917123, + 0.0017930219182744622, + 0.0024192919954657555, + 0.002953135408461094, + 0.010251732543110847, + 0.0036507986951619387, + 0.005638308823108673, + 0.006688771769404411, + 0.004589044488966465, + 0.002789664315059781, + 0.00299538834951818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.02059001475572586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "It's an extreme close-up, shot with a still camera in realistic, high-speed slow motion, focusing on a lightweight paper receipt resting on a dark, polished granite kitchen countertop. Made of thin, off-white thermal paper, the receipt is slightly wrinkled, with sharp creases and lines of text printed in faded, dark ink visible but barely discernible. A gentle, unseen breeze, originating off-camera, begins cautiously lifting one edge of the paper. The receipt rises in exquisite slow motion, undulating in an elegant, almost liquid dance, its thin body twisting and fluttering inches above the smooth surface, revealing every subtle ripple and twist as it momentarily hovers in mid-air. As the air currents subside, the receipt begins to descend, deliberately floating down, landing back on the countertop, not flat but with a soft, final curl, one corner touching first, before the rest of the paper gently collapses. Soft morning light streams across the scene from a nearby window, capturing the fluttering edges of the paper with bright highlights and casting long, soft, dynamic shadows that move with every movement of the receipt, creating a serene, contemplative, and visually mesmerizing atmosphere, completely focused on the delicate physics of this ordinary moment.", + "seed": 22565, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_11/index_5422_seed_22565_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0843303062953055, + "vlm_logprob": -0.0843303062953055, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.0843303062953055, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07663551717996597 + ], + "opposite_logprob_tokens": [ + -2.7016355991363525 + ], + "logprob_reward": -0.07663551717996597, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07663551717996597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41496339440345764, + 0.07923277467489243, + 0.04708588123321533, + 0.054563067853450775, + 0.02954099513590336, + 0.21305130422115326, + 0.08939353376626968, + 0.020166045054793358, + 0.031737156212329865, + 0.02026587538421154 + ], + "credit_map": [ + 0.07793103158473969, + 0.03894132375717163, + 0.16281159222126007, + 0.030619051307439804, + 0.018566828221082687, + 0.03604717180132866, + 0.013080156408250332, + 0.036966271698474884, + 0.006589415017515421, + 0.009607182815670967, + 0.007358587812632322, + 0.009044095873832703, + 0.011327570304274559, + 0.010529291816055775, + 0.012364559806883335, + 0.01241206843405962, + 0.003784480271860957, + 0.006276232656091452, + 0.004252179060131311, + 0.0034889844246208668, + 0.010478883050382137, + 0.005715776700526476, + 0.008607866242527962, + 0.004481480456888676, + 0.004163730423897505, + 0.008214614354074001, + 0.005820380989462137, + 0.004971412010490894, + 0.009814552031457424, + 0.007754268124699593, + 0.008009509183466434, + 0.005814601667225361, + 0.0020258789882063866, + 0.005475257057696581, + 0.0025352961383759975, + 0.002472918014973402, + 0.008708958514034748, + 0.004025663714855909, + 0.0020846298430114985, + 0.0022123921662569046, + 0.0054252189584076405, + 0.0882801041007042, + 0.003918475937098265, + 0.0024782363325357437, + 0.09905914217233658, + 0.010809418745338917, + 0.0015787769807502627, + 0.0015019432175904512, + 0.00564078614115715, + 0.01988619938492775, + 0.007281824946403503, + 0.006010229699313641, + 0.026234397664666176, + 0.01176898181438446, + 0.006785816513001919, + 0.005785302259027958, + 0.0012856473913416266, + 0.0022746077738702297, + 0.0016144539695233107, + 0.00259060924872756, + 0.002958992961794138, + 0.005248512607067823, + 0.0020276065915822983, + 0.002165614627301693, + 0.0029631485231220722, + 0.003914146218448877, + 0.004043650347739458, + 0.00529819680377841, + 0.0031859935261309147, + 0.004933794494718313, + 0.003751939395442605, + 0.0036462850403040648, + 0.001156429061666131, + 0.002930478425696492, + 0.0018093072576448321, + 0.0024723089300096035, + 0.004021866712719202, + 0.004008842166513205, + 0.0020888186991214752, + 0.0017778241308405995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.07663551717996597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293431848287582 + ], + "opposite_logprob_tokens": [ + -3.7732934951782227 + ], + "logprob_reward": -0.023293431848287582, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293431848287582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5435619354248047, + 0.03520406410098076, + 0.05392575263977051, + 0.06193454936146736, + 0.06442367285490036, + 0.08488241583108902, + 0.06833839416503906, + 0.012817607261240482, + 0.04816504940390587, + 0.026746487244963646 + ], + "credit_map": [ + 0.09557081758975983, + 0.044057607650756836, + 0.18273305892944336, + 0.036914926022291183, + 0.03018829971551895, + 0.053201913833618164, + 0.033180251717567444, + 0.06771505624055862, + 0.0034133943263441324, + 0.004860639106482267, + 0.002893290249630809, + 0.00393613800406456, + 0.0039004948921501637, + 0.0066169206984341145, + 0.005174135323613882, + 0.004409050568938255, + 0.004170470871031284, + 0.006241425406187773, + 0.0072132740169763565, + 0.0056940726935863495, + 0.007312274072319269, + 0.009805595502257347, + 0.007177591789513826, + 0.006311051081866026, + 0.0065130628645420074, + 0.009797526523470879, + 0.007248353213071823, + 0.004798943642526865, + 0.013918429613113403, + 0.009034691378474236, + 0.005146266892552376, + 0.00547727569937706, + 0.003271585563197732, + 0.020671406760811806, + 0.00295030465349555, + 0.0036508708726614714, + 0.02088652178645134, + 0.00696119898930192, + 0.002514228457584977, + 0.003517564618960023, + 0.008902566507458687, + 0.014854070730507374, + 0.008999226614832878, + 0.008241624571383, + 0.016392283141613007, + 0.018244776874780655, + 0.0047658211551606655, + 0.004482049960643053, + 0.004466771613806486, + 0.013907945714890957, + 0.004021069500595331, + 0.0036502378061413765, + 0.01889295130968094, + 0.011424791067838669, + 0.006301832851022482, + 0.00567279476672411, + 0.0010518166236579418, + 0.0016772497911006212, + 0.0014569431077688932, + 0.0017399538774043322, + 0.0020609700586646795, + 0.0021422659046947956, + 0.0012829032493755221, + 0.0014055046485736966, + 0.0032796585001051426, + 0.00865557324141264, + 0.00516123604029417, + 0.008079027757048607, + 0.00832100585103035, + 0.005917795468121767, + 0.005116554442793131, + 0.0036341974046081305, + 0.0023369768168777227, + 0.003953888546675444, + 0.0036949727218598127, + 0.0038973609916865826, + 0.0032141192350536585, + 0.0046049426309764385, + 0.0030509845819324255, + 0.001993240090087056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.023293431848287582 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31347155570983887 + ], + "opposite_logprob_tokens": [ + -1.3134715557098389 + ], + "logprob_reward": -0.31347155570983887, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31347155570983887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31772908568382263, + 0.034427110105752945, + 0.038017790764570236, + 0.0458449050784111, + 0.104408860206604, + 0.25200355052948, + 0.057818204164505005, + 0.09184761345386505, + 0.040594711899757385, + 0.01730818673968315 + ], + "credit_map": [ + 0.052493251860141754, + 0.030689598992466927, + 0.11377785354852676, + 0.02363506332039833, + 0.023216169327497482, + 0.028872661292552948, + 0.01817058026790619, + 0.02687389776110649, + 0.0034748343750834465, + 0.004506511613726616, + 0.004222815856337547, + 0.003685966832563281, + 0.005188779905438423, + 0.00358950556255877, + 0.0050516072660684586, + 0.0047070905566215515, + 0.00400152150541544, + 0.0051174527034163475, + 0.00404807785525918, + 0.0029338246677070856, + 0.006792474538087845, + 0.004357333295047283, + 0.006543318275362253, + 0.0042237876914441586, + 0.004635905381292105, + 0.006665072403848171, + 0.004866064991801977, + 0.00485497061163187, + 0.0070266094990074635, + 0.005869656801223755, + 0.00732877803966403, + 0.004597845487296581, + 0.0038650301285088062, + 0.03591727465391159, + 0.00416515814140439, + 0.0034346438478678465, + 0.04193226620554924, + 0.007280249614268541, + 0.0034826386254280806, + 0.004331601317971945, + 0.009134966880083084, + 0.0921948179602623, + 0.00869237631559372, + 0.007568287197500467, + 0.10408160090446472, + 0.021306956186890602, + 0.004610132426023483, + 0.004414417315274477, + 0.0036119818687438965, + 0.010539304465055466, + 0.0037385905161499977, + 0.005712546408176422, + 0.01383813563734293, + 0.010964637622237206, + 0.00476615596562624, + 0.004646847024559975, + 0.0074258712120354176, + 0.018712367862462997, + 0.006920226383954287, + 0.007241500541567802, + 0.020139414817094803, + 0.016759004443883896, + 0.007494198624044657, + 0.00715503329411149, + 0.002572093391790986, + 0.0063778432086110115, + 0.004115753341466188, + 0.005627546925097704, + 0.005827095825225115, + 0.006169953383505344, + 0.005659686401486397, + 0.004244741052389145, + 0.001346126664429903, + 0.001868000952526927, + 0.0026228532660752535, + 0.003214182797819376, + 0.0025029510725289583, + 0.0026485936250537634, + 0.0016415818827226758, + 0.0014638954307883978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.31347155570983887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015150746330618858 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015150746330618858, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015150746330618858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3851991891860962, + 0.021889682859182358, + 0.0544322207570076, + 0.05294881388545036, + 0.1250452697277069, + 0.1866069883108139, + 0.061288002878427505, + 0.04471089690923691, + 0.04330284148454666, + 0.02457609958946705 + ], + "credit_map": [ + 0.05190828815102577, + 0.03302350640296936, + 0.1429978609085083, + 0.029840752482414246, + 0.027105120941996574, + 0.0476679801940918, + 0.019266700372099876, + 0.03338899835944176, + 0.00180804671254009, + 0.002495993161574006, + 0.0021915941033512354, + 0.002926051616668701, + 0.003827761160209775, + 0.002928943606093526, + 0.0033713411539793015, + 0.0023399521596729755, + 0.004568890202790499, + 0.005939680617302656, + 0.007065750192850828, + 0.005636921152472496, + 0.006019641645252705, + 0.010349649004638195, + 0.0070351362228393555, + 0.007816547527909279, + 0.006190148647874594, + 0.008943583816289902, + 0.0030580400489270687, + 0.003458951599895954, + 0.012473927810788155, + 0.006856646854430437, + 0.006085016764700413, + 0.005882502533495426, + 0.004443175159394741, + 0.04774292930960655, + 0.00353219429962337, + 0.003817847464233637, + 0.05212390422821045, + 0.006413981784135103, + 0.0034769754856824875, + 0.0034942589700222015, + 0.0058877673000097275, + 0.0760592669248581, + 0.004167415201663971, + 0.002747613936662674, + 0.08401282876729965, + 0.009808841161429882, + 0.002117878058925271, + 0.0018053761450573802, + 0.004092584364116192, + 0.012621281668543816, + 0.003716699779033661, + 0.0032669533975422382, + 0.016770081594586372, + 0.00986325740814209, + 0.005765155889093876, + 0.00519199064001441, + 0.0036611228715628386, + 0.00814664177596569, + 0.003508242778480053, + 0.004197548143565655, + 0.007998425513505936, + 0.009249389171600342, + 0.004111772403120995, + 0.0038377561140805483, + 0.003356505651026964, + 0.007641931530088186, + 0.004294374492019415, + 0.006349267438054085, + 0.007710236590355635, + 0.005195703823119402, + 0.005012274254113436, + 0.0037425507325679064, + 0.0016243407735601068, + 0.0037795270327478647, + 0.0020676401909440756, + 0.0026876835618168116, + 0.005179226864129305, + 0.004451058804988861, + 0.0024160065222531557, + 0.002370616188272834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0015150746330618858 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735952105373144 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735952105373144, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735952105373144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.085169218480587, + 0.03351360186934471, + 0.024722591042518616, + 0.05712109059095383, + 0.07234742492437363, + 0.3068031966686249, + 0.06062792241573334, + 0.18197722733020782, + 0.05304878577589989, + 0.12466893345117569 + ], + "credit_map": [ + 0.012624367140233517, + 0.00907963514328003, + 0.010254932567477226, + 0.005998462904244661, + 0.013407144695520401, + 0.01495668850839138, + 0.011128641664981842, + 0.007719343062490225, + 0.0031069545075297356, + 0.0037757870741188526, + 0.002429461106657982, + 0.004141515586525202, + 0.004840558394789696, + 0.006228458136320114, + 0.005216097459197044, + 0.00377477309666574, + 0.00337897683493793, + 0.003201458603143692, + 0.0028722158167511225, + 0.002051197923719883, + 0.003574143862351775, + 0.0040917834267020226, + 0.0033808057196438313, + 0.00217200955376029, + 0.00652013486251235, + 0.008769280277192593, + 0.004046017769724131, + 0.004279688000679016, + 0.01074609998613596, + 0.009502243250608444, + 0.004880781751126051, + 0.00837684515863657, + 0.006148411892354488, + 0.013277918100357056, + 0.008100125007331371, + 0.004879062995314598, + 0.013729488477110863, + 0.015660826116800308, + 0.004839315544813871, + 0.0057122791185975075, + 0.010385925881564617, + 0.12321435660123825, + 0.0068815541453659534, + 0.005102187395095825, + 0.13660040497779846, + 0.01789860986173153, + 0.0032269058283418417, + 0.003493261057883501, + 0.0048082685098052025, + 0.011238034814596176, + 0.003950844518840313, + 0.004061263054609299, + 0.014188447035849094, + 0.011784575879573822, + 0.005356920417398214, + 0.005239568650722504, + 0.01501005794852972, + 0.03695352375507355, + 0.01415910106152296, + 0.014418265782296658, + 0.03768929839134216, + 0.036278024315834045, + 0.014197004027664661, + 0.013271939009428024, + 0.00449188519269228, + 0.007331571076065302, + 0.00487812515348196, + 0.005351397208869457, + 0.0075737470760941505, + 0.014369748532772064, + 0.004190994426608086, + 0.004861318506300449, + 0.011594211682677269, + 0.01647058315575123, + 0.018739130347967148, + 0.021352585405111313, + 0.014296761713922024, + 0.021407848224043846, + 0.011741569265723228, + 0.009066241793334484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.006735952105373144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "It's an extreme close-up, shot with a still camera in realistic, high-speed slow motion, focusing on a lightweight paper receipt resting on a dark, polished granite kitchen countertop. Made of thin, off-white thermal paper, the receipt is slightly wrinkled, with sharp creases and lines of text printed in faded, dark ink visible but barely discernible. A gentle, unseen breeze, originating off-camera, begins cautiously lifting one edge of the paper. The receipt rises in exquisite slow motion, undulating in an elegant, almost liquid dance, its thin body twisting and fluttering inches above the smooth surface, revealing every subtle ripple and twist as it momentarily hovers in mid-air. As the air currents subside, the receipt begins to descend, deliberately floating down, landing back on the countertop, not flat but with a soft, final curl, one corner touching first, before the rest of the paper gently collapses. Soft morning light streams across the scene from a nearby window, capturing the fluttering edges of the paper with bright highlights and casting long, soft, dynamic shadows that move with every movement of the receipt, creating a serene, contemplative, and visually mesmerizing atmosphere, completely focused on the delicate physics of this ordinary moment.", + "seed": 22566, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_11/index_5422_seed_22566_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2759528153575957, + "vlm_logprob": -0.2759528153575957, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.2759528153575957, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09853588789701462 + ], + "opposite_logprob_tokens": [ + -2.4735357761383057 + ], + "logprob_reward": -0.09853588789701462, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09853588789701462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4855574071407318, + 0.054272085428237915, + 0.02933685854077339, + 0.19453787803649902, + 0.03412211686372757, + 0.026620494201779366, + 0.09057370573282242, + 0.03382095322012901, + 0.034834880381822586, + 0.016323627904057503 + ], + "credit_map": [ + 0.03685711324214935, + 0.015049954876303673, + 0.15401406586170197, + 0.017343631014227867, + 0.014558680355548859, + 0.04246571287512779, + 0.02246146649122238, + 0.18280678987503052, + 0.004676968790590763, + 0.004362593404948711, + 0.0057174451649188995, + 0.008710759691894054, + 0.004974589683115482, + 0.00760191585868597, + 0.009095587767660618, + 0.009132223203778267, + 0.0025314034428447485, + 0.002952615497633815, + 0.0037370279897004366, + 0.003384922631084919, + 0.004582767840474844, + 0.0043646953999996185, + 0.003928233869373798, + 0.003855191869661212, + 0.011925233528017998, + 0.01653463765978813, + 0.03697190061211586, + 0.024539334699511528, + 0.021282529458403587, + 0.03825967386364937, + 0.025019465014338493, + 0.02000509947538376, + 0.0023110981564968824, + 0.0069684418849647045, + 0.006078705657273531, + 0.0027302061207592487, + 0.004941893741488457, + 0.005055218935012817, + 0.003281861310824752, + 0.0027546926867216825, + 0.002142895944416523, + 0.004086812026798725, + 0.004324297420680523, + 0.0020203967578709126, + 0.0038898729253560305, + 0.0038001269567757845, + 0.002260762033984065, + 0.004095329437404871, + 0.006377760320901871, + 0.007185978349298239, + 0.02342933416366577, + 0.004089053254574537, + 0.015270504169166088, + 0.023183340206742287, + 0.006090512964874506, + 0.004947221837937832, + 0.0027186230290681124, + 0.00428680470213294, + 0.004882179666310549, + 0.0035394132137298584, + 0.00543223787099123, + 0.004674870055168867, + 0.004225343000143766, + 0.004061480984091759, + 0.003472311655059457, + 0.005839576944708824, + 0.005196513142436743, + 0.003055274486541748, + 0.005453627090901136, + 0.006448670756071806, + 0.002088639186695218, + 0.003280265023931861, + 0.0013241852866485715, + 0.0016189208254218102, + 0.004396019037812948, + 0.0008622604655101895, + 0.0020195667166262865, + 0.004583891946822405, + 0.0007887078681960702, + 0.0007300766301341355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.09853588789701462 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811738267540932 + ], + "opposite_logprob_tokens": [ + -3.2881174087524414 + ], + "logprob_reward": -0.03811738267540932, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811738267540932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7697203755378723, + 0.04511592909693718, + 0.03145073726773262, + 0.04390672966837883, + 0.03383972868323326, + 0.013869533315300941, + 0.013880203478038311, + 0.020193375647068024, + 0.017112135887145996, + 0.010911266319453716 + ], + "credit_map": [ + 0.05726759135723114, + 0.02291574329137802, + 0.24431806802749634, + 0.027404263615608215, + 0.01981101930141449, + 0.07439018040895462, + 0.02697908692061901, + 0.29663440585136414, + 0.0045316885225474834, + 0.004594302270561457, + 0.006662280764430761, + 0.005711235571652651, + 0.004808810539543629, + 0.005414215382188559, + 0.0055402349680662155, + 0.007853157818317413, + 0.003656385699287057, + 0.004488098900765181, + 0.004251033533364534, + 0.0035704979673027992, + 0.0031453617848455906, + 0.0047188326716423035, + 0.0030060147400945425, + 0.004614511039108038, + 0.0028823697939515114, + 0.003892500652000308, + 0.00834282673895359, + 0.005271583329886198, + 0.0051094042137265205, + 0.008850674144923687, + 0.005362901836633682, + 0.004194469656795263, + 0.002808765508234501, + 0.006249851081520319, + 0.006335168611258268, + 0.0022768359631299973, + 0.005893102381378412, + 0.005990490783005953, + 0.0019013312412425876, + 0.00238418229855597, + 0.0012652003206312656, + 0.0021300558000802994, + 0.0016086548566818237, + 0.0012597653549164534, + 0.0026302789337933064, + 0.0016123902751132846, + 0.0015712432796135545, + 0.0017919447273015976, + 0.0011507533490657806, + 0.0020210985094308853, + 0.0014045053394511342, + 0.0014353494625538588, + 0.0021998737938702106, + 0.002392648020759225, + 0.0015082312747836113, + 0.0017677440773695707, + 0.0014932332560420036, + 0.0020381673239171505, + 0.005356136243790388, + 0.0015774390194565058, + 0.0028001933824270964, + 0.004053483717143536, + 0.0014759866753593087, + 0.0013987347483634949, + 0.0018263441743329167, + 0.002366088330745697, + 0.0035631651990115643, + 0.0011413503671064973, + 0.0027349626179784536, + 0.0031812344677746296, + 0.0012394660152494907, + 0.001059525879099965, + 0.0007526152185164392, + 0.0009199328487738967, + 0.0024790854658931494, + 0.0006580513436347246, + 0.0022075194865465164, + 0.0027541581075638533, + 0.0006619703490287066, + 0.000477933936053887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.03811738267540932 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236438989639282 + ], + "opposite_logprob_tokens": [ + -0.34864386916160583 + ], + "logprob_reward": -1.2236438989639282, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236438989639282, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4692637622356415, + 0.03326679766178131, + 0.054284509271383286, + 0.03970056772232056, + 0.16424444317817688, + 0.02222103253006935, + 0.04030158370733261, + 0.07170099765062332, + 0.09574927389621735, + 0.009267089888453484 + ], + "credit_map": [ + 0.028615562245249748, + 0.011229872703552246, + 0.04574821516871452, + 0.011061026714742184, + 0.013923805207014084, + 0.1450832635164261, + 0.019032688811421394, + 0.19456925988197327, + 0.003372452687472105, + 0.0034879788290709257, + 0.004687776323407888, + 0.004334806930273771, + 0.0033851051703095436, + 0.004562388174235821, + 0.00415672454982996, + 0.005279561039060354, + 0.005757031962275505, + 0.0068749673664569855, + 0.00676006730645895, + 0.006847770418971777, + 0.006600911729037762, + 0.01038809958845377, + 0.005118611268699169, + 0.005937043111771345, + 0.0033030835911631584, + 0.004183289594948292, + 0.004406750202178955, + 0.005194703117012978, + 0.004911061841994524, + 0.004838571883738041, + 0.008149667643010616, + 0.004713436122983694, + 0.011034791357815266, + 0.03905609995126724, + 0.02938065677881241, + 0.01205139234662056, + 0.03183584660291672, + 0.021124163642525673, + 0.008990702219307423, + 0.010770785622298717, + 0.0021614658180624247, + 0.0025456936564296484, + 0.0024707247503101826, + 0.001853382564149797, + 0.0028090791311115026, + 0.004615549463778734, + 0.0021188671234995127, + 0.0036462671123445034, + 0.003609859850257635, + 0.00486514950171113, + 0.007221762556582689, + 0.003469536080956459, + 0.00624417420476675, + 0.007604099810123444, + 0.0036696502938866615, + 0.003617348149418831, + 0.005298730451613665, + 0.009942722506821156, + 0.016679223626852036, + 0.005364147014915943, + 0.013220981694757938, + 0.011899825185537338, + 0.005035002250224352, + 0.004260359797626734, + 0.0066916681826114655, + 0.009426177479326725, + 0.0256490521132946, + 0.007039360702037811, + 0.014016380533576012, + 0.021925179287791252, + 0.005336109548807144, + 0.005665336735546589, + 0.0007151099853217602, + 0.0009404275333508849, + 0.00229586660861969, + 0.0007882923819124699, + 0.001308061764575541, + 0.001931502716615796, + 0.0007099462090991437, + 0.0005778825143352151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -1.2236438989639282 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005261146929115057 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.005261146929115057, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261146929115057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4591306746006012, + 0.05810533091425896, + 0.04369198903441429, + 0.05909818410873413, + 0.10190632939338684, + 0.05087298899888992, + 0.06419479101896286, + 0.059106115251779556, + 0.042046528309583664, + 0.06184700131416321 + ], + "credit_map": [ + 0.02599233016371727, + 0.014331115409731865, + 0.03801406919956207, + 0.019873807206749916, + 0.022964948788285255, + 0.16952520608901978, + 0.023502424359321594, + 0.14492680132389069, + 0.005695409141480923, + 0.007294976152479649, + 0.006568687502294779, + 0.006795400287955999, + 0.006657722871750593, + 0.007251399103552103, + 0.00854343082755804, + 0.009298311546444893, + 0.003999671898782253, + 0.004961458966135979, + 0.005551171954721212, + 0.0050422390922904015, + 0.005136555060744286, + 0.008346215821802616, + 0.004989175125956535, + 0.005665506701916456, + 0.00516679510474205, + 0.007760656997561455, + 0.006572728045284748, + 0.007544815540313721, + 0.007921534590423107, + 0.007402356248348951, + 0.008598603308200836, + 0.008130702190101147, + 0.006554353516548872, + 0.021997831761837006, + 0.015930304303765297, + 0.008923479355871677, + 0.01624993607401848, + 0.011946678161621094, + 0.01029421016573906, + 0.010009543038904667, + 0.004917588550597429, + 0.004936841782182455, + 0.008525010198354721, + 0.004414074122905731, + 0.00783666130155325, + 0.007043944671750069, + 0.004469667561352253, + 0.008729206398129463, + 0.00630338815972209, + 0.01175585575401783, + 0.010197775438427925, + 0.0033734117168933153, + 0.012296078726649284, + 0.009926464408636093, + 0.005006411578506231, + 0.005335413385182619, + 0.004112935625016689, + 0.0070381853729486465, + 0.010070854797959328, + 0.0045944442972540855, + 0.01499138306826353, + 0.007535808719694614, + 0.005607925355434418, + 0.005154586397111416, + 0.004039873834699392, + 0.004747370257973671, + 0.005774967838078737, + 0.005929199978709221, + 0.005925020668655634, + 0.006173767149448395, + 0.004971159156411886, + 0.004485175013542175, + 0.0035278566647320986, + 0.007135188207030296, + 0.011777856387197971, + 0.004467197693884373, + 0.01849254220724106, + 0.008447858504951, + 0.004641321487724781, + 0.0033571829553693533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.005261146929115057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lightweight paper receipt made of thin, off-white thermal paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the receipt resting on a dark, polished granite kitchen countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the video shot with a still camera in realistic, high-speed slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the receipt rise, undulate, twist, and flutter inches above the surface before descending and landing with a soft curl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft morning light streaming from a nearby window, creating bright highlights and long, soft, dynamic shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205760322511196 + ], + "opposite_logprob_tokens": [ + -4.2642059326171875 + ], + "logprob_reward": -0.014205760322511196, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205760322511196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3142833709716797, + 0.07042615860700607, + 0.09352648258209229, + 0.03849034756422043, + 0.15102015435695648, + 0.07262906432151794, + 0.06263388693332672, + 0.04658877104520798, + 0.09003866463899612, + 0.06036309897899628 + ], + "credit_map": [ + 0.018724123015999794, + 0.011565238237380981, + 0.09698419272899628, + 0.019930578768253326, + 0.018888963386416435, + 0.09358279407024384, + 0.020793236792087555, + 0.033814240247011185, + 0.006291598081588745, + 0.0058100102469325066, + 0.011104190722107887, + 0.008285876363515854, + 0.007070692256093025, + 0.009851982817053795, + 0.011683345772325993, + 0.010328459553420544, + 0.009490177035331726, + 0.010732743889093399, + 0.008601764217019081, + 0.01237319316715002, + 0.011267364025115967, + 0.01447196863591671, + 0.012932262383401394, + 0.01365701388567686, + 0.0039382074028253555, + 0.004739955998957157, + 0.005399980582296848, + 0.0028351026121526957, + 0.0058126202784478664, + 0.007115022279322147, + 0.004311495926231146, + 0.00433796551078558, + 0.01344999298453331, + 0.02240542694926262, + 0.024980224668979645, + 0.009550170972943306, + 0.024286245927214622, + 0.033925559371709824, + 0.008795898407697678, + 0.013626644387841225, + 0.005631563253700733, + 0.012799817137420177, + 0.018204763531684875, + 0.004599041771143675, + 0.01051523070782423, + 0.010903802700340748, + 0.00426087249070406, + 0.005713972728699446, + 0.004555224906653166, + 0.009973262436687946, + 0.013376756571233273, + 0.004987266380339861, + 0.01081897784024477, + 0.010516876354813576, + 0.002930557122454047, + 0.005474965553730726, + 0.0046852161176502705, + 0.006182656157761812, + 0.008606772869825363, + 0.004612883552908897, + 0.0077159591019153595, + 0.006507046055048704, + 0.003933920990675688, + 0.004344316199421883, + 0.007628986611962318, + 0.010267595760524273, + 0.021473297849297523, + 0.005757799372076988, + 0.014416594989597797, + 0.019160686060786247, + 0.00573993194848299, + 0.005593766458332539, + 0.0065307714976370335, + 0.010388242080807686, + 0.008349666371941566, + 0.00451979786157608, + 0.013172611594200134, + 0.0081211868673563, + 0.005121618043631315, + 0.004159206990152597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.014205760322511196 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_12/index_tensor([1849, 1849])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_12/index_tensor([1849, 1849])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2588e6d08789d4bcff77da1f1094919477d356e6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_12/index_tensor([1849, 1849])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows a vintage steam train as it races through a deep, precipitous valley. The train's jet-black metal, stained with age-worn rust and oil, its massive red drive wheels and intricate linkage mechanism churn at a frantic pace, roaring along tracks that gleam coldly in the morning light. From its towering chimneys, thick, billowing plumes of white steam erupt violently into the sky. The golden morning sun pierces the mist, tinging the magnificent clouds with a warm golden edge. The steam is then quickly torn away by the powerful air currents of the train's passing. Flanked by steep slopes covered in lush, dark green pine forests, the train's entire body trembles slightly at high speed, creating a grand, imposing scene imbued with a sense of power, speed, and the unstoppable grandeur of the industrial age.", + "seed": 1097, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_12/index_1849_seed_1097_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.028868029359728098, + "vlm_logprob": -0.028868029359728098, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.028868029359728098, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017174800857901573 + ], + "opposite_logprob_tokens": [ + -4.39217472076416 + ], + "logprob_reward": -0.017174800857901573, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.017174800857901573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0855962336063385, + 0.023596974089741707, + 0.022249065339565277, + 0.04314122721552849, + 0.028377901762723923, + 0.02673838846385479, + 0.048761963844299316, + 0.20802107453346252, + 0.45257848501205444, + 0.06093868613243103 + ], + "credit_map": [ + 0.01296220812946558, + 0.01064305566251278, + 0.008907432667911053, + 0.010585875250399113, + 0.014833910390734673, + 0.008736631833016872, + 0.011356199160218239, + 0.007570919580757618, + 0.002947277622297406, + 0.002294476144015789, + 0.0026677006389945745, + 0.002787903882563114, + 0.004043184220790863, + 0.0025847272481769323, + 0.003350061597302556, + 0.002921642968431115, + 0.002723266137763858, + 0.0028326655738055706, + 0.003202969441190362, + 0.0031648706644773483, + 0.0037238316144794226, + 0.0024497006088495255, + 0.0020410234574228525, + 0.0021107380744069815, + 0.006069259252399206, + 0.0036473830696195364, + 0.008645394816994667, + 0.009694159962236881, + 0.0051392558962106705, + 0.004380146507173777, + 0.0026353751309216022, + 0.0029302509501576424, + 0.0031627377029508352, + 0.0028880725149065256, + 0.004165684338659048, + 0.005583322141319513, + 0.0025200233794748783, + 0.002436948474496603, + 0.004711282439529896, + 0.0029098307713866234, + 0.0023122369311749935, + 0.003460551146417856, + 0.003347082296386361, + 0.0036493621300905943, + 0.0026394466403871775, + 0.003138182917609811, + 0.005464041605591774, + 0.0027274852618575096, + 0.006370754446834326, + 0.008238065987825394, + 0.007363570388406515, + 0.006807718891650438, + 0.005663147196173668, + 0.006043192930519581, + 0.0043582250364124775, + 0.003917287103831768, + 0.02204926684498787, + 0.04522157460451126, + 0.037246719002723694, + 0.02795572392642498, + 0.023497790098190308, + 0.013228379189968109, + 0.022398658096790314, + 0.01642296276986599, + 0.040314435958862305, + 0.02026473358273506, + 0.06432829797267914, + 0.06915415078401566, + 0.03025089204311371, + 0.07175419479608536, + 0.12477442622184753, + 0.03173735365271568, + 0.005220542196184397, + 0.009670339524745941, + 0.010808109305799007, + 0.013206083327531815, + 0.00556282140314579, + 0.00375173008069396, + 0.007043189834803343, + 0.005675869062542915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.017174800857901573 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865693673491478 + ], + "opposite_logprob_tokens": [ + -3.048656940460205 + ], + "logprob_reward": -0.04865693673491478, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865693673491478, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15719768404960632, + 0.15437045693397522, + 0.23044632375240326, + 0.1705458164215088, + 0.02434290573000908, + 0.01806090585887432, + 0.05237693712115288, + 0.07971447706222534, + 0.06015392392873764, + 0.05279048904776573 + ], + "credit_map": [ + 0.022655392065644264, + 0.023257536813616753, + 0.020243778824806213, + 0.018734978511929512, + 0.020096711814403534, + 0.016734154894948006, + 0.017884360626339912, + 0.017590779811143875, + 0.02307157963514328, + 0.028481461107730865, + 0.02430194802582264, + 0.023222941905260086, + 0.02200194261968136, + 0.015648989006876945, + 0.008939221501350403, + 0.00870237685739994, + 0.024352453649044037, + 0.01870930753648281, + 0.02498168684542179, + 0.0272463858127594, + 0.05866246297955513, + 0.0313679501414299, + 0.027958808466792107, + 0.017167283222079277, + 0.021655306220054626, + 0.014414800330996513, + 0.028715096414089203, + 0.03563831374049187, + 0.020573822781443596, + 0.016121041029691696, + 0.017642252147197723, + 0.015785187482833862, + 0.0024202389176934958, + 0.002868923358619213, + 0.004716919735074043, + 0.004461795091629028, + 0.0018920085858553648, + 0.0022731965873390436, + 0.003082149662077427, + 0.0026276735588908195, + 0.0013977345079183578, + 0.0020431189332157373, + 0.0036125765182077885, + 0.0025347834452986717, + 0.0015806989977136254, + 0.0019113257294520736, + 0.0033501707948744297, + 0.0016304984455928206, + 0.005458645988255739, + 0.00854080356657505, + 0.009674625471234322, + 0.006034005433320999, + 0.005894818808883429, + 0.006855502724647522, + 0.005686556454747915, + 0.004231979604810476, + 0.00816228799521923, + 0.014398310333490372, + 0.006166795268654823, + 0.009525509551167488, + 0.015529334545135498, + 0.00632841233164072, + 0.011978533118963242, + 0.007625292520970106, + 0.005615255329757929, + 0.002414831193163991, + 0.006997063755989075, + 0.0072156088426709175, + 0.0035151089541614056, + 0.011064853519201279, + 0.01983005180954933, + 0.0035011430736631155, + 0.004514731466770172, + 0.004476100672036409, + 0.010258358903229237, + 0.015942925587296486, + 0.0027331106830388308, + 0.004149510990828276, + 0.0075794062577188015, + 0.003136346349492669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.04865693673491478 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067374915815889835 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.0067374915815889835, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067374915815889835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11915843933820724, + 0.12624269723892212, + 0.07896056026220322, + 0.03449324890971184, + 0.021435968577861786, + 0.02244068682193756, + 0.09060987085103989, + 0.18389739096164703, + 0.255283385515213, + 0.06747779995203018 + ], + "credit_map": [ + 0.019542299211025238, + 0.01676291972398758, + 0.013052009977400303, + 0.017356710508465767, + 0.016892991960048676, + 0.010306403040885925, + 0.014545516110956669, + 0.010699592530727386, + 0.028424082323908806, + 0.02373781055212021, + 0.024558838456869125, + 0.015435601584613323, + 0.01214691437780857, + 0.009169580414891243, + 0.007302987389266491, + 0.005466881673783064, + 0.009038328193128109, + 0.0052062952890992165, + 0.007277032360434532, + 0.007416315842419863, + 0.019802602007985115, + 0.012844693847000599, + 0.011040687561035156, + 0.006334605161100626, + 0.005237201694399118, + 0.0025577314663678408, + 0.007127402350306511, + 0.010330033488571644, + 0.0024823343846946955, + 0.002687783446162939, + 0.002230887534096837, + 0.001839877455495298, + 0.002462539356201887, + 0.0015696502523496747, + 0.004554255399852991, + 0.0037251177709549665, + 0.0017942117992788553, + 0.0018686471739783883, + 0.0030390333849936724, + 0.0024225134402513504, + 0.0023452513851225376, + 0.0036391555331647396, + 0.003965064883232117, + 0.0036437285598367453, + 0.0022046579979360104, + 0.0018969279481098056, + 0.002633612370118499, + 0.00211228895932436, + 0.00808858685195446, + 0.010021797381341457, + 0.01683623343706131, + 0.011811534874141216, + 0.01177388709038496, + 0.013581284321844578, + 0.009301546961069107, + 0.009195004589855671, + 0.018805380910634995, + 0.03890559449791908, + 0.01744932308793068, + 0.02847885526716709, + 0.024670476093888283, + 0.012707407586276531, + 0.02241065353155136, + 0.020469699054956436, + 0.02347647212445736, + 0.010905380360782146, + 0.03299565613269806, + 0.03312281146645546, + 0.016316251829266548, + 0.044369932264089584, + 0.07713650166988373, + 0.016960371285676956, + 0.005318836774677038, + 0.009552126750349998, + 0.010278075002133846, + 0.01406947337090969, + 0.011471038684248924, + 0.004811751656234264, + 0.007216417230665684, + 0.004760073963552713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.0067374915815889835 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368956968188286 + ], + "opposite_logprob_tokens": [ + -3.408689498901367 + ], + "logprob_reward": -0.03368956968188286, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368956968188286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08148117363452911, + 0.17164506018161774, + 0.034903634339571, + 0.1663643717765808, + 0.0333205983042717, + 0.014769433066248894, + 0.18049564957618713, + 0.07817896455526352, + 0.18755024671554565, + 0.05129086598753929 + ], + "credit_map": [ + 0.012365496717393398, + 0.009467902593314648, + 0.01019515935331583, + 0.013610602356493473, + 0.011705001816153526, + 0.00692529184743762, + 0.00998952891677618, + 0.007222190964967012, + 0.03527246415615082, + 0.026676135137677193, + 0.028686631470918655, + 0.02613290213048458, + 0.01697300560772419, + 0.012662767432630062, + 0.013952775858342648, + 0.011288371868431568, + 0.0037074005231261253, + 0.00317657389678061, + 0.004433704074472189, + 0.0050229234620928764, + 0.007693230174481869, + 0.004191295243799686, + 0.00419159559533, + 0.002486912067979574, + 0.02326403371989727, + 0.012259643524885178, + 0.028395039960741997, + 0.039249710738658905, + 0.018088605254888535, + 0.020439283922314644, + 0.012731646187603474, + 0.011936414055526257, + 0.003398048924282193, + 0.0033166715875267982, + 0.006455678027123213, + 0.005155073944479227, + 0.003225978696718812, + 0.002864865120500326, + 0.004959965590387583, + 0.00394431222230196, + 0.001093242084607482, + 0.0016590538434684277, + 0.0031856498681008816, + 0.002110104076564312, + 0.0012973009143024683, + 0.0012430677888914943, + 0.0024718360509723425, + 0.0017091779736801982, + 0.016029957681894302, + 0.015566237270832062, + 0.03815387934446335, + 0.0177407618612051, + 0.028702273964881897, + 0.033441245555877686, + 0.015052154660224915, + 0.015809141099452972, + 0.008064853958785534, + 0.015083860605955124, + 0.011500677093863487, + 0.011803042143583298, + 0.011392357759177685, + 0.007057584822177887, + 0.007296604570001364, + 0.005979984067380428, + 0.01641976274549961, + 0.00866060983389616, + 0.028597792610526085, + 0.031630080193281174, + 0.01204727590084076, + 0.027798384428024292, + 0.049451861530542374, + 0.012944484129548073, + 0.0035398888867348433, + 0.004836597945541143, + 0.00788062997162342, + 0.014161305502057076, + 0.006284035742282867, + 0.0031796602997928858, + 0.008418763987720013, + 0.00298998411744833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.03368956968188286 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038081347942352295 + ], + "opposite_logprob_tokens": [ + -3.288081407546997 + ], + "logprob_reward": -0.038081347942352295, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038081347942352295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09237375110387802, + 0.03988683223724365, + 0.23431384563446045, + 0.11079105734825134, + 0.021978119388222694, + 0.017546793445944786, + 0.04037172347307205, + 0.03588136285543442, + 0.1476966142654419, + 0.2591599225997925 + ], + "credit_map": [ + 0.013413302600383759, + 0.015089555643498898, + 0.009517449885606766, + 0.012830903753638268, + 0.013612821698188782, + 0.009143202565610409, + 0.010379079729318619, + 0.008387436158955097, + 0.005718956235796213, + 0.005912963300943375, + 0.005728626623749733, + 0.006347188726067543, + 0.006039285100996494, + 0.0036017769016325474, + 0.0033186753280460835, + 0.003219361649826169, + 0.03184792026877403, + 0.017203502357006073, + 0.030430244281888008, + 0.022198792546987534, + 0.05804098770022392, + 0.03131327033042908, + 0.02948520891368389, + 0.013793899677693844, + 0.02003050409257412, + 0.007599084638059139, + 0.030584728345274925, + 0.028096484020352364, + 0.005225616507232189, + 0.006949462462216616, + 0.006749109365046024, + 0.005556062329560518, + 0.0025399327278137207, + 0.0018984833732247353, + 0.004233118612319231, + 0.004947307053953409, + 0.0017755922162905335, + 0.0014497973024845123, + 0.002936725504696369, + 0.002197162015363574, + 0.0018339678645133972, + 0.0017283449415117502, + 0.003975554835051298, + 0.002723611192777753, + 0.0016539947828277946, + 0.001438854611478746, + 0.00240136100910604, + 0.0017911047907546163, + 0.00466828653588891, + 0.0062270136550068855, + 0.005421470385044813, + 0.005718294996768236, + 0.004773629829287529, + 0.003537979209795594, + 0.005324878264218569, + 0.004700170364230871, + 0.0036353578325361013, + 0.0034955774899572134, + 0.003693106584250927, + 0.005567011423408985, + 0.005910991225391626, + 0.002233251929283142, + 0.005261854734271765, + 0.006084210705012083, + 0.012669732794165611, + 0.007391405291855335, + 0.023179378360509872, + 0.025285108014941216, + 0.009097269736230373, + 0.02156262844800949, + 0.03811538219451904, + 0.010395714081823826, + 0.015061612240970135, + 0.020393867045640945, + 0.04322014003992081, + 0.08208315074443817, + 0.01958654262125492, + 0.01284983940422535, + 0.054461926221847534, + 0.011502841487526894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.038081347942352295 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a vintage steam train as it races through a deep, precipitous valley. The train's jet-black metal, stained with age-worn rust and oil, its massive red drive wheels and intricate linkage mechanism churn at a frantic pace, roaring along tracks that gleam coldly in the morning light. From its towering chimneys, thick, billowing plumes of white steam erupt violently into the sky. The golden morning sun pierces the mist, tinging the magnificent clouds with a warm golden edge. The steam is then quickly torn away by the powerful air currents of the train's passing. Flanked by steep slopes covered in lush, dark green pine forests, the train's entire body trembles slightly at high speed, creating a grand, imposing scene imbued with a sense of power, speed, and the unstoppable grandeur of the industrial age.", + "seed": 1098, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_12/index_1849_seed_1098_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024348144605755805, + "vlm_logprob": -0.024348144605755805, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.024348144605755805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012264786288142204 + ], + "opposite_logprob_tokens": [ + -4.762264728546143 + ], + "logprob_reward": -0.012264786288142204, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012264786288142204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08245222270488739, + 0.5329552292823792, + 0.03181101009249687, + 0.18411296606063843, + 0.028561128303408623, + 0.017477115616202354, + 0.05457880347967148, + 0.01548496913164854, + 0.020334890112280846, + 0.032231591641902924 + ], + "credit_map": [ + 0.008063387125730515, + 0.009344431571662426, + 0.009543066844344139, + 0.016168931499123573, + 0.013010912574827671, + 0.011796307750046253, + 0.006937314290553331, + 0.007587891072034836, + 0.036422133445739746, + 0.02844782918691635, + 0.016365094110369682, + 0.02098849043250084, + 0.20124392211437225, + 0.19299881160259247, + 0.02172301895916462, + 0.014765985310077667, + 0.0024431529454886913, + 0.002458720002323389, + 0.002915083197876811, + 0.0036214650608599186, + 0.004367878194898367, + 0.007362821605056524, + 0.005072820466011763, + 0.0035690749064087868, + 0.019694646820425987, + 0.01374860294163227, + 0.015199047513306141, + 0.032669056206941605, + 0.02765963412821293, + 0.025582673028111458, + 0.03328830003738403, + 0.016271043568849564, + 0.004776096437126398, + 0.0030532770324498415, + 0.0030068992637097836, + 0.003407663432881236, + 0.003305493388324976, + 0.0031488570384681225, + 0.004499707370996475, + 0.0033631380647420883, + 0.0018851123750209808, + 0.002096229698508978, + 0.0022343697492033243, + 0.0023169887717813253, + 0.0018817975651472807, + 0.0017716448055580258, + 0.0024301509838551283, + 0.0028608241118490696, + 0.011508949100971222, + 0.005811986513435841, + 0.008774464018642902, + 0.005440606269985437, + 0.004858341533690691, + 0.0037245044950395823, + 0.009420569986104965, + 0.005039395298808813, + 0.0018035124521702528, + 0.0010084244422614574, + 0.002235598396509886, + 0.002230620477348566, + 0.001366734504699707, + 0.0011337746400386095, + 0.0020604573655873537, + 0.0036458512768149376, + 0.002016540616750717, + 0.0021474312525242567, + 0.003622213378548622, + 0.0035240494180470705, + 0.0023676673881709576, + 0.0020921416580677032, + 0.002229712437838316, + 0.00233513955026865, + 0.005486705806106329, + 0.003959614783525467, + 0.004554304294288158, + 0.004124506376683712, + 0.0028137811459600925, + 0.004360244143754244, + 0.0034440108574926853, + 0.0034884295891970396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.012264786288142204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865739122033119 + ], + "opposite_logprob_tokens": [ + -3.0486574172973633 + ], + "logprob_reward": -0.04865739122033119, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865739122033119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1678493618965149, + 0.31663528084754944, + 0.03380164876580238, + 0.1329326629638672, + 0.0323338508605957, + 0.07756195962429047, + 0.10114922374486923, + 0.039167776703834534, + 0.034033920615911484, + 0.06453432887792587 + ], + "credit_map": [ + 0.013777154497802258, + 0.016988571733236313, + 0.016896583139896393, + 0.03655329719185829, + 0.01870029605925083, + 0.025091640651226044, + 0.01655840501189232, + 0.023283401504158974, + 0.028647420927882195, + 0.017136871814727783, + 0.019455743953585625, + 0.028625385835766792, + 0.07373548299074173, + 0.0874791070818901, + 0.03854003921151161, + 0.02301519550383091, + 0.0034550426062196493, + 0.005000711884349585, + 0.002228308469057083, + 0.003886338323354721, + 0.005091279745101929, + 0.005353962536901236, + 0.005099966656416655, + 0.0036860371474176645, + 0.01382450107485056, + 0.009382475167512894, + 0.012578976340591908, + 0.02319089137017727, + 0.017450803890824318, + 0.01718517206609249, + 0.02714724838733673, + 0.012172579765319824, + 0.004883986432105303, + 0.005026364233344793, + 0.004814813379198313, + 0.003971563186496496, + 0.0034890053793787956, + 0.0030112883541733027, + 0.0036473535001277924, + 0.003489475231617689, + 0.009690859355032444, + 0.015377305448055267, + 0.010298051871359348, + 0.009701239876449108, + 0.007386903744190931, + 0.005295031704008579, + 0.009846669621765614, + 0.009965892881155014, + 0.02177259512245655, + 0.008682528510689735, + 0.02541792392730713, + 0.009001153521239758, + 0.00508606992661953, + 0.005956787616014481, + 0.016436859965324402, + 0.00879529770463705, + 0.005543693900108337, + 0.004204799421131611, + 0.004945066757500172, + 0.004923994652926922, + 0.004250390455126762, + 0.002905698958784342, + 0.005735983606427908, + 0.0066581470891833305, + 0.0026124194264411926, + 0.0029360714834183455, + 0.007589292712509632, + 0.005165420472621918, + 0.003115409519523382, + 0.004028641153126955, + 0.004722265992313623, + 0.0038643970619887114, + 0.007173873949795961, + 0.005478935316205025, + 0.011572044342756271, + 0.0046616531908512115, + 0.0043790265917778015, + 0.007828665897250175, + 0.007369308732450008, + 0.01607082039117813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.04865739122033119 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107686385512352 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.01107686385512352, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107686385512352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16346216201782227, + 0.5564458966255188, + 0.0957866758108139, + 0.019583292305469513, + 0.01329880952835083, + 0.012239976786077023, + 0.08008935302495956, + 0.019758837297558784, + 0.009884602390229702, + 0.029450461268424988 + ], + "credit_map": [ + 0.013278110884130001, + 0.011378355324268341, + 0.008770378306508064, + 0.022852355614304543, + 0.02134537510573864, + 0.04787835478782654, + 0.02174784243106842, + 0.016211388632655144, + 0.04224665090441704, + 0.012556072324514389, + 0.01008985098451376, + 0.01020007487386465, + 0.22192882001399994, + 0.22272159159183502, + 0.01842574030160904, + 0.01827702485024929, + 0.011348379775881767, + 0.009503867477178574, + 0.0074106077663600445, + 0.010483424179255962, + 0.014704786241054535, + 0.01560176257044077, + 0.014812466688454151, + 0.011921376921236515, + 0.0022408596705645323, + 0.0019285025773569942, + 0.0012440283317118883, + 0.0021756943315267563, + 0.003608198370784521, + 0.0030296053737401962, + 0.003548078704625368, + 0.0018083250615745783, + 0.002337204059585929, + 0.002064968692138791, + 0.0018842186545953155, + 0.0012413433287292719, + 0.0012630517594516277, + 0.0012624580413103104, + 0.001832241890951991, + 0.0014133229851722717, + 0.0013924838276579976, + 0.0015486631309613585, + 0.0013050136622041464, + 0.0015901012811809778, + 0.0011153175728395581, + 0.0010061984648928046, + 0.0019207957666367292, + 0.002361402614042163, + 0.014081358909606934, + 0.006364523898810148, + 0.010911500081419945, + 0.00857857521623373, + 0.007044636178761721, + 0.0055710445158183575, + 0.018873879685997963, + 0.008663838729262352, + 0.0028720307163894176, + 0.001732871518470347, + 0.003136578481644392, + 0.0028080528136342764, + 0.0018395116785541177, + 0.001153846038505435, + 0.0026844851672649384, + 0.003531459951773286, + 0.0009146761731244624, + 0.0011397978523746133, + 0.001214688061736524, + 0.001474318327382207, + 0.0014078589156270027, + 0.001107043121010065, + 0.001516012824140489, + 0.001110207405872643, + 0.002333806362003088, + 0.0038661558646708727, + 0.003983579576015472, + 0.004340678453445435, + 0.0024198449682444334, + 0.0040563722141087055, + 0.004494622815400362, + 0.003955401014536619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01107686385512352 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061563044786453 + ], + "opposite_logprob_tokens": [ + -4.141061782836914 + ], + "logprob_reward": -0.016061563044786453, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061563044786453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14024126529693604, + 0.4136563837528229, + 0.12040165811777115, + 0.0745239108800888, + 0.021262075752019882, + 0.07121967524290085, + 0.026535093784332275, + 0.017749404534697533, + 0.009311338886618614, + 0.10509917140007019 + ], + "credit_map": [ + 0.013829573057591915, + 0.009544650092720985, + 0.006945083383470774, + 0.01610439270734787, + 0.01918409764766693, + 0.04250209406018257, + 0.01696324162185192, + 0.015168135985732079, + 0.029455017298460007, + 0.007557782344520092, + 0.007983619347214699, + 0.010675973258912563, + 0.1621159315109253, + 0.16191543638706207, + 0.01847698912024498, + 0.015475614927709103, + 0.015741607174277306, + 0.017405839636921883, + 0.009417641907930374, + 0.014354908838868141, + 0.016962788999080658, + 0.01895650289952755, + 0.016674373298883438, + 0.010887996293604374, + 0.008610386401414871, + 0.00455102464184165, + 0.007138436194509268, + 0.012934460304677486, + 0.007780957035720348, + 0.012587436474859715, + 0.012753213755786419, + 0.008167993277311325, + 0.003148826537653804, + 0.002923101419582963, + 0.002857785439118743, + 0.0030759191140532494, + 0.0019931297283619642, + 0.0024443957954645157, + 0.0023116993252187967, + 0.00250721862539649, + 0.008713997900485992, + 0.014820411801338196, + 0.011282410472631454, + 0.007694723084568977, + 0.006604334339499474, + 0.005520742852240801, + 0.00721560325473547, + 0.009367448277771473, + 0.0073114545084536076, + 0.0023916158825159073, + 0.007656023371964693, + 0.0022883121855556965, + 0.001697562518529594, + 0.001404418027959764, + 0.0025729998014867306, + 0.0012127070222049952, + 0.0022002910263836384, + 0.0023468295112252235, + 0.002516099950298667, + 0.0020929821766912937, + 0.002284323563799262, + 0.0013171747559681535, + 0.002370235975831747, + 0.002621467225253582, + 0.0010204447899013758, + 0.0006560772890225053, + 0.0009235402103513479, + 0.0014905661810189486, + 0.0009996384615078568, + 0.0009962175972759724, + 0.0021171767730265856, + 0.0011076774680987, + 0.024036308750510216, + 0.012042718008160591, + 0.01813322864472866, + 0.00874254945665598, + 0.007435740437358618, + 0.010792568325996399, + 0.008585448376834393, + 0.015330610796809196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.016061563044786453 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368011862039566 + ], + "opposite_logprob_tokens": [ + -3.4086802005767822 + ], + "logprob_reward": -0.03368011862039566, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368011862039566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07732656598091125, + 0.5761522054672241, + 0.07141813635826111, + 0.02358555607497692, + 0.020980335772037506, + 0.020985618233680725, + 0.11327171325683594, + 0.014389565214514732, + 0.013256986625492573, + 0.0686333104968071 + ], + "credit_map": [ + 0.0072437613271176815, + 0.0058974167332053185, + 0.005645187571644783, + 0.010606259107589722, + 0.011586618609726429, + 0.013715192675590515, + 0.011669667437672615, + 0.010962463915348053, + 0.04240037128329277, + 0.011010144837200642, + 0.01189475692808628, + 0.010172044858336449, + 0.2387091964483261, + 0.23128902912139893, + 0.014401117339730263, + 0.016275569796562195, + 0.0059458352625370026, + 0.007695364765822887, + 0.004400169476866722, + 0.007013258989900351, + 0.009456592611968517, + 0.015039937570691109, + 0.014103432185947895, + 0.007763545494526625, + 0.0033493267837911844, + 0.0025138496421277523, + 0.0016493741422891617, + 0.003053649328649044, + 0.003888670587912202, + 0.0033558730501681566, + 0.003493763506412506, + 0.0022810494992882013, + 0.003519886638969183, + 0.002630512462928891, + 0.00260904710739851, + 0.0021560927852988243, + 0.002555521670728922, + 0.002190333092585206, + 0.0032187567558139563, + 0.002100184094160795, + 0.0024257260374724865, + 0.00258831400424242, + 0.0027492723893374205, + 0.003669154131785035, + 0.0023195084650069475, + 0.001300412812270224, + 0.003592010820284486, + 0.002341220621019602, + 0.02528592012822628, + 0.011242012493312359, + 0.020356107503175735, + 0.010680954903364182, + 0.007138085085898638, + 0.0066489530727267265, + 0.022087402641773224, + 0.009832275100052357, + 0.00180195237044245, + 0.0014131198404356837, + 0.0016583956312388182, + 0.0019651686307042837, + 0.0010306383483111858, + 0.0008641732856631279, + 0.0016342269955202937, + 0.004021889064460993, + 0.0009915418922901154, + 0.0016107677947729826, + 0.0023398695047944784, + 0.002443839330226183, + 0.0013231405755504966, + 0.0011281525949016213, + 0.0019523967057466507, + 0.0014672776451334357, + 0.01334103662520647, + 0.005494652781635523, + 0.010467277839779854, + 0.005495918449014425, + 0.004206253681331873, + 0.0073073431849479675, + 0.007043114863336086, + 0.015277707949280739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.03368011862039566 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a vintage steam train as it races through a deep, precipitous valley. The train's jet-black metal, stained with age-worn rust and oil, its massive red drive wheels and intricate linkage mechanism churn at a frantic pace, roaring along tracks that gleam coldly in the morning light. From its towering chimneys, thick, billowing plumes of white steam erupt violently into the sky. The golden morning sun pierces the mist, tinging the magnificent clouds with a warm golden edge. The steam is then quickly torn away by the powerful air currents of the train's passing. Flanked by steep slopes covered in lush, dark green pine forests, the train's entire body trembles slightly at high speed, creating a grand, imposing scene imbued with a sense of power, speed, and the unstoppable grandeur of the industrial age.", + "seed": 1099, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_12/index_1849_seed_1099_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02695939764380455, + "vlm_logprob": -0.02695939764380455, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.02695939764380455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017396362498402596 + ], + "opposite_logprob_tokens": [ + -4.267396450042725 + ], + "logprob_reward": -0.017396362498402596, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017396362498402596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06691473722457886, + 0.0972432792186737, + 0.02785930037498474, + 0.20535050332546234, + 0.0569782480597496, + 0.11039463430643082, + 0.1778465211391449, + 0.13051097095012665, + 0.03390646353363991, + 0.09299540519714355 + ], + "credit_map": [ + 0.007028575520962477, + 0.012063289992511272, + 0.005608800332993269, + 0.008933760225772858, + 0.010097112506628036, + 0.00836576335132122, + 0.00840351078659296, + 0.0064139170572161674, + 0.005333124194294214, + 0.021059047430753708, + 0.004275480750948191, + 0.010536041110754013, + 0.0311863012611866, + 0.010563607327640057, + 0.008523478172719479, + 0.005766195710748434, + 0.0042222668416798115, + 0.004644237924367189, + 0.0021938029676675797, + 0.003968123346567154, + 0.004673201125115156, + 0.0028763492591679096, + 0.002555299550294876, + 0.0027260170318186283, + 0.01880401372909546, + 0.02644566260278225, + 0.021196983754634857, + 0.03550480678677559, + 0.02941700629889965, + 0.017228152602910995, + 0.035429976880550385, + 0.02132388949394226, + 0.005967982579022646, + 0.00424926308915019, + 0.003387185512110591, + 0.007346002850681543, + 0.005891844630241394, + 0.016355806961655617, + 0.007997837848961353, + 0.005782318767160177, + 0.016608675941824913, + 0.010890549048781395, + 0.012698880396783352, + 0.015286275185644627, + 0.010193136520683765, + 0.010707567445933819, + 0.016092553734779358, + 0.017916996031999588, + 0.01455414667725563, + 0.018871108070015907, + 0.014098059386014938, + 0.022808607667684555, + 0.015742948278784752, + 0.019272418692708015, + 0.040470998734235764, + 0.03202822431921959, + 0.010608294047415257, + 0.009193032048642635, + 0.017366528511047363, + 0.018356435000896454, + 0.01100881677120924, + 0.014861409552395344, + 0.01876208744943142, + 0.030354352667927742, + 0.004213941749185324, + 0.0033345308620482683, + 0.0036652677226811647, + 0.0024787585716694593, + 0.0034697488881647587, + 0.0052709272131323814, + 0.0037883503828197718, + 0.007684934418648481, + 0.006881921552121639, + 0.018458474427461624, + 0.00890597328543663, + 0.008650976233184338, + 0.02150443010032177, + 0.011188461445271969, + 0.008099972270429134, + 0.009305184707045555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.017396362498402596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026352232322096825 + ], + "opposite_logprob_tokens": [ + -3.6513521671295166 + ], + "logprob_reward": -0.026352232322096825, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026352232322096825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1603761613368988, + 0.06986239552497864, + 0.09056071192026138, + 0.08171624690294266, + 0.1294487565755844, + 0.10827893763780594, + 0.10775057971477509, + 0.11322841793298721, + 0.0605945959687233, + 0.0781831219792366 + ], + "credit_map": [ + 0.016505541279911995, + 0.021374354138970375, + 0.017738454043865204, + 0.02033349871635437, + 0.01513036247342825, + 0.02106841467320919, + 0.026013385504484177, + 0.022212164476513863, + 0.0055798860266804695, + 0.012579040601849556, + 0.006901502143591642, + 0.007732881233096123, + 0.017657283693552017, + 0.007481924258172512, + 0.005540021229535341, + 0.0063898591324687, + 0.010865308344364166, + 0.010194729082286358, + 0.013561660423874855, + 0.016210507601499557, + 0.008196054957807064, + 0.013441630639135838, + 0.008453106507658958, + 0.009637720882892609, + 0.010239455848932266, + 0.016598591580986977, + 0.007030955515801907, + 0.009932456538081169, + 0.014090396463871002, + 0.0068923779763281345, + 0.009495021775364876, + 0.007437004707753658, + 0.013377546332776546, + 0.013722428120672703, + 0.00878568273037672, + 0.016107704490423203, + 0.010824675671756268, + 0.039905086159706116, + 0.01497227419167757, + 0.011753367260098457, + 0.015204964205622673, + 0.010033237747848034, + 0.012156463228166103, + 0.014462212100625038, + 0.01241709291934967, + 0.02135412022471428, + 0.008597282692790031, + 0.014053587801754475, + 0.010111246258020401, + 0.015526833012700081, + 0.009773959405720234, + 0.014170889742672443, + 0.014369872398674488, + 0.018813680857419968, + 0.011603167280554771, + 0.013380948454141617, + 0.014199486002326012, + 0.008346755057573318, + 0.01422968227416277, + 0.026340866461396217, + 0.0074819098226726055, + 0.013157513923943043, + 0.013636325486004353, + 0.015835890546441078, + 0.007014150731265545, + 0.008224284276366234, + 0.0072601959109306335, + 0.00634244317188859, + 0.006420411169528961, + 0.007874579168856144, + 0.007159721106290817, + 0.010298817418515682, + 0.008370045572519302, + 0.01597808301448822, + 0.006913705263286829, + 0.006246991455554962, + 0.018059762194752693, + 0.007424680516123772, + 0.006764170713722706, + 0.008425693958997726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.026352232322096825 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160688366740942 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160688366740942, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160688366740942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06564547121524811, + 0.47537073493003845, + 0.11104056984186172, + 0.12497962266206741, + 0.07163046300411224, + 0.040075767785310745, + 0.018050868064165115, + 0.03512754663825035, + 0.02319362200796604, + 0.034885305911302567 + ], + "credit_map": [ + 0.006811133120208979, + 0.011202598921954632, + 0.005508203525096178, + 0.00951892789453268, + 0.009787194430828094, + 0.0077939750626683235, + 0.008475401438772678, + 0.006548034027218819, + 0.024545397609472275, + 0.10344304889440536, + 0.020571794360876083, + 0.0526948980987072, + 0.15182355046272278, + 0.05005798488855362, + 0.04481050372123718, + 0.02742358297109604, + 0.020966786891222, + 0.021656310185790062, + 0.007459329906851053, + 0.011498196050524712, + 0.0191313698887825, + 0.01120640616863966, + 0.009482037276029587, + 0.009640134871006012, + 0.015565942041575909, + 0.026296840980648994, + 0.01497096847742796, + 0.014354628510773182, + 0.017140062525868416, + 0.014295042492449284, + 0.012074032798409462, + 0.01028210949152708, + 0.00787275843322277, + 0.006916619371622801, + 0.003808720735833049, + 0.008544227108359337, + 0.006023421883583069, + 0.022578753530979156, + 0.007751588709652424, + 0.008134374395012856, + 0.004862650763243437, + 0.0029477067291736603, + 0.004020106513053179, + 0.005731220822781324, + 0.00479576550424099, + 0.009512857533991337, + 0.0038558519445359707, + 0.004349608905613422, + 0.0017048290465027094, + 0.002504756674170494, + 0.0018132495461031795, + 0.002037959871813655, + 0.0027007334865629673, + 0.003818662604317069, + 0.0016578180948272347, + 0.0018128573428839445, + 0.004354191478341818, + 0.0038160455878823996, + 0.005047298502177, + 0.004993191454559565, + 0.0032308015506714582, + 0.005852089263498783, + 0.003956366330385208, + 0.0038775629363954067, + 0.001999122090637684, + 0.0031702262349426746, + 0.0023320664186030626, + 0.0019884638022631407, + 0.0027228922117501497, + 0.004449141211807728, + 0.0027358061634004116, + 0.003795904340222478, + 0.0039017514791339636, + 0.00629466725513339, + 0.003434109967201948, + 0.002873844699934125, + 0.0063996696844697, + 0.003347870660945773, + 0.004136050119996071, + 0.004497339017689228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.0160688366740942 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864683002233505 + ], + "opposite_logprob_tokens": [ + -3.048646926879883 + ], + "logprob_reward": -0.04864683002233505, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864683002233505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22825491428375244, + 0.3295043408870697, + 0.06503075361251831, + 0.16434283554553986, + 0.09350956231355667, + 0.019608335569500923, + 0.016165239736437798, + 0.030519140884280205, + 0.029050312936306, + 0.024014601483941078 + ], + "credit_map": [ + 0.013483335264027119, + 0.04441610723733902, + 0.019123630598187447, + 0.025244999676942825, + 0.05449502542614937, + 0.01590290106832981, + 0.03515646234154701, + 0.020432475954294205, + 0.01679430529475212, + 0.07214785367250443, + 0.013529544696211815, + 0.0361030139029026, + 0.10463880747556686, + 0.03538854420185089, + 0.031353868544101715, + 0.019548406824469566, + 0.017784615978598595, + 0.01630713976919651, + 0.0032502091489732265, + 0.0038862917572259903, + 0.01010163128376007, + 0.00557833444327116, + 0.004177320282906294, + 0.0039452132768929005, + 0.020349781960248947, + 0.03377537801861763, + 0.021375879645347595, + 0.019263913854956627, + 0.02374875359237194, + 0.014661517925560474, + 0.018689613789319992, + 0.012478005141019821, + 0.01033884659409523, + 0.009388691745698452, + 0.006118310149759054, + 0.011244645342230797, + 0.010161970742046833, + 0.02762775495648384, + 0.01072941068559885, + 0.007899941876530647, + 0.0019034456927329302, + 0.0024670034181326628, + 0.0030983195174485445, + 0.0027331807650625706, + 0.0025712696369737387, + 0.0036534040700644255, + 0.001269018859602511, + 0.0019126960542052984, + 0.0015933358808979392, + 0.0025266753509640694, + 0.001953703351318836, + 0.00200423551723361, + 0.0024089389480650425, + 0.0026367977261543274, + 0.0014081676490604877, + 0.0016333876410499215, + 0.003062012605369091, + 0.0025506082456558943, + 0.006028397474437952, + 0.0032975042704492807, + 0.002783867996186018, + 0.005628990475088358, + 0.0041081407107412815, + 0.0030596223659813404, + 0.0026697067078202963, + 0.0035698902793228626, + 0.0034678508527576923, + 0.002658989280462265, + 0.0030071239452809095, + 0.00392050389200449, + 0.004140235483646393, + 0.005616014823317528, + 0.0024958299472928047, + 0.0027932943776249886, + 0.0025781660806387663, + 0.0025945729576051235, + 0.005057965870946646, + 0.0027473128866404295, + 0.0022419148590415716, + 0.003505547298118472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.04864683002233505 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026332726702094078 + ], + "opposite_logprob_tokens": [ + -3.6513326168060303 + ], + "logprob_reward": -0.026332726702094078, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026332726702094078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09276460111141205, + 0.0714370533823967, + 0.018968427553772926, + 0.11187012493610382, + 0.14027826488018036, + 0.12341898679733276, + 0.06220078840851784, + 0.16164211928844452, + 0.12309246510267258, + 0.09432706981897354 + ], + "credit_map": [ + 0.010825960896909237, + 0.015270848758518696, + 0.008016983047127724, + 0.01251534279435873, + 0.015488561242818832, + 0.006134881172329187, + 0.012920842505991459, + 0.011591184884309769, + 0.004047810565680265, + 0.01539677195250988, + 0.003044417593628168, + 0.007421365939080715, + 0.02243444323539734, + 0.007604672107845545, + 0.006899174302816391, + 0.0045883990824222565, + 0.002240446861833334, + 0.0026659946888685226, + 0.0020942094270139933, + 0.0024347524158656597, + 0.0035159129183739424, + 0.0019195298664271832, + 0.0024588690139353275, + 0.0016387113137170672, + 0.010452727787196636, + 0.010586611926555634, + 0.010962060652673244, + 0.017632154747843742, + 0.01299200113862753, + 0.00935034453868866, + 0.023063773289322853, + 0.016830449923872948, + 0.015274448320269585, + 0.009733765386044979, + 0.00811680592596531, + 0.018456444144248962, + 0.010882508009672165, + 0.045038167387247086, + 0.019168369472026825, + 0.013607766479253769, + 0.01577470637857914, + 0.011348609812557697, + 0.015658412128686905, + 0.014538316056132317, + 0.014316924847662449, + 0.028284849599003792, + 0.0100455516949296, + 0.013451613485813141, + 0.008353549987077713, + 0.009203699417412281, + 0.00785251334309578, + 0.007521273102611303, + 0.0066681066527962685, + 0.007774360943585634, + 0.008298752829432487, + 0.006528535392135382, + 0.019777419045567513, + 0.019064154475927353, + 0.02021172270178795, + 0.027177510783076286, + 0.018000593408942223, + 0.026275906711816788, + 0.017132820561528206, + 0.014001995325088501, + 0.008744246326386929, + 0.01996636763215065, + 0.01633347012102604, + 0.013921658508479595, + 0.01799999363720417, + 0.01460583508014679, + 0.011576289311051369, + 0.019944606348872185, + 0.00946995709091425, + 0.01598905399441719, + 0.009301193989813328, + 0.008714789524674416, + 0.018473075702786446, + 0.010878418572247028, + 0.010601907037198544, + 0.010898680426180363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.026332726702094078 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a vintage steam train as it races through a deep, precipitous valley. The train's jet-black metal, stained with age-worn rust and oil, its massive red drive wheels and intricate linkage mechanism churn at a frantic pace, roaring along tracks that gleam coldly in the morning light. From its towering chimneys, thick, billowing plumes of white steam erupt violently into the sky. The golden morning sun pierces the mist, tinging the magnificent clouds with a warm golden edge. The steam is then quickly torn away by the powerful air currents of the train's passing. Flanked by steep slopes covered in lush, dark green pine forests, the train's entire body trembles slightly at high speed, creating a grand, imposing scene imbued with a sense of power, speed, and the unstoppable grandeur of the industrial age.", + "seed": 1100, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_12/index_1849_seed_1100_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04013730511069298, + "vlm_logprob": -0.04013730511069298, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.04013730511069298, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02412801794707775 + ], + "opposite_logprob_tokens": [ + -4.024127960205078 + ], + "logprob_reward": -0.02412801794707775, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02412801794707775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06885583698749542, + 0.07636359333992004, + 0.24306415021419525, + 0.0654296875, + 0.11046181619167328, + 0.21455954015254974, + 0.06996527314186096, + 0.04503707215189934, + 0.034940462559461594, + 0.07132253795862198 + ], + "credit_map": [ + 0.00960858166217804, + 0.010771730914711952, + 0.00637769466266036, + 0.011683781631290913, + 0.006736056413501501, + 0.0043306732550263405, + 0.011862228624522686, + 0.007485087029635906, + 0.012062134221196175, + 0.012627214193344116, + 0.006624555680900812, + 0.007290778215974569, + 0.01040685549378395, + 0.008854140527546406, + 0.010359356179833412, + 0.008138560689985752, + 0.02901551127433777, + 0.0439591184258461, + 0.030308540910482407, + 0.0218935776501894, + 0.041306354105472565, + 0.026034779846668243, + 0.025137070566415787, + 0.025409212335944176, + 0.007097181864082813, + 0.010958426631987095, + 0.008057489059865475, + 0.003935330547392368, + 0.005752487573772669, + 0.010021129623055458, + 0.006335767917335033, + 0.013271872885525227, + 0.01571446843445301, + 0.012198947370052338, + 0.00963661726564169, + 0.006337402854114771, + 0.016671888530254364, + 0.020341461524367332, + 0.013897553086280823, + 0.015663480386137962, + 0.022417927160859108, + 0.0249194148927927, + 0.018298598006367683, + 0.021771084517240524, + 0.03616988658905029, + 0.037154942750930786, + 0.028027690947055817, + 0.025799985975027084, + 0.009212218225002289, + 0.01511556375771761, + 0.005528060719370842, + 0.007210317999124527, + 0.009856237098574638, + 0.009360828436911106, + 0.006558602210134268, + 0.007123452145606279, + 0.004249953664839268, + 0.0065023028291761875, + 0.004708342254161835, + 0.004624921828508377, + 0.008324583061039448, + 0.00503978505730629, + 0.004769083112478256, + 0.006818098481744528, + 0.0036203269846737385, + 0.004097662400454283, + 0.005397387780249119, + 0.004691894166171551, + 0.00353340245783329, + 0.005238456651568413, + 0.0035103217232972383, + 0.00485101155936718, + 0.007887659594416618, + 0.007753782905638218, + 0.008444703184068203, + 0.008326662704348564, + 0.008471700362861156, + 0.012677308171987534, + 0.008639877662062645, + 0.009120840579271317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.02412801794707775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495959892868996 + ], + "opposite_logprob_tokens": [ + -2.929959535598755 + ], + "logprob_reward": -0.05495959892868996, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495959892868996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20707851648330688, + 0.0462985523045063, + 0.07889994978904724, + 0.28931862115859985, + 0.05132367089390755, + 0.14691931009292603, + 0.033867768943309784, + 0.043609946966171265, + 0.027078982442617416, + 0.0756046324968338 + ], + "credit_map": [ + 0.03588150441646576, + 0.03609393537044525, + 0.040640395134687424, + 0.021818235516548157, + 0.020412394776940346, + 0.01599065773189068, + 0.017363166436553, + 0.01887824386358261, + 0.008527534082531929, + 0.008273863233625889, + 0.0056410240940749645, + 0.003731992095708847, + 0.006213689688593149, + 0.0028713075444102287, + 0.007404847536236048, + 0.003634294494986534, + 0.008832750841975212, + 0.009895184077322483, + 0.006701716221868992, + 0.006003376096487045, + 0.009337597526609898, + 0.012101257219910622, + 0.012131164781749249, + 0.013896911405026913, + 0.050379678606987, + 0.04155224934220314, + 0.041110146790742874, + 0.020855160430073738, + 0.017628008499741554, + 0.05358897149562836, + 0.018906865268945694, + 0.04529755190014839, + 0.005563173908740282, + 0.00659432215616107, + 0.005788808688521385, + 0.005000090226531029, + 0.0068703447468578815, + 0.005733039695769548, + 0.008975180797278881, + 0.006798716727644205, + 0.014754197560250759, + 0.019464900717139244, + 0.012524046003818512, + 0.018907617777585983, + 0.020800556987524033, + 0.02372989058494568, + 0.018839923664927483, + 0.017898187041282654, + 0.004458641167730093, + 0.0045100850984454155, + 0.004235965199768543, + 0.00341854989528656, + 0.004863882437348366, + 0.004655075259506702, + 0.0038492009043693542, + 0.0038763717748224735, + 0.0060114143416285515, + 0.011235535144805908, + 0.0033893820364028215, + 0.00540043693035841, + 0.005021106451749802, + 0.0031157727353274822, + 0.00518815778195858, + 0.004248148296028376, + 0.0022948882542550564, + 0.0036773609463125467, + 0.0040459828451275826, + 0.003277421463280916, + 0.004233836196362972, + 0.00405730027705431, + 0.0024750709999352694, + 0.0030171216931194067, + 0.009392846375703812, + 0.012120692990720272, + 0.012581647373735905, + 0.008094705641269684, + 0.0082528255879879, + 0.009398887865245342, + 0.007251514121890068, + 0.008511519059538841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.05495959892868996 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463874451816082 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.00463874451816082, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463874451816082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08507151901721954, + 0.20200008153915405, + 0.16107088327407837, + 0.04881151765584946, + 0.09373984485864639, + 0.15263813734054565, + 0.04618091881275177, + 0.1325823813676834, + 0.029342912137508392, + 0.048561856150627136 + ], + "credit_map": [ + 0.009309202432632446, + 0.010532869957387447, + 0.016557374969124794, + 0.013221320696175098, + 0.008825687691569328, + 0.007871459238231182, + 0.00927756354212761, + 0.009476032108068466, + 0.032581210136413574, + 0.0301173385232687, + 0.02673451043665409, + 0.018628863617777824, + 0.025614339858293533, + 0.019504711031913757, + 0.02116047404706478, + 0.027658607810735703, + 0.017244068905711174, + 0.018993260338902473, + 0.017360234633088112, + 0.014428346417844296, + 0.022399896755814552, + 0.019027097150683403, + 0.027042290195822716, + 0.024575665593147278, + 0.007093924563378096, + 0.007401518989354372, + 0.006009064149111509, + 0.0037944670766592026, + 0.004425958730280399, + 0.007618892006576061, + 0.006221492774784565, + 0.006246194709092379, + 0.010342276655137539, + 0.013252433389425278, + 0.010843097232282162, + 0.006503351032733917, + 0.014400510117411613, + 0.012874367646872997, + 0.015323139727115631, + 0.01020065974444151, + 0.018576273694634438, + 0.019557924941182137, + 0.013497128151357174, + 0.01304110698401928, + 0.023952007293701172, + 0.028827331960201263, + 0.016265438869595528, + 0.018920907750725746, + 0.007283169310539961, + 0.0058869849890470505, + 0.004340553656220436, + 0.0046311067417263985, + 0.005865192040801048, + 0.008049760945141315, + 0.004301225766539574, + 0.005822920240461826, + 0.016996579244732857, + 0.03313595801591873, + 0.013153457082808018, + 0.012315020896494389, + 0.018208468332886696, + 0.012354287318885326, + 0.011068289168179035, + 0.015350311994552612, + 0.0025071785785257816, + 0.0032514154445379972, + 0.00376801285892725, + 0.0031835855916142464, + 0.005760230123996735, + 0.004476627334952354, + 0.0024572545662522316, + 0.0039386022835969925, + 0.0051268949173390865, + 0.005649061873555183, + 0.0064626419916749, + 0.005594294983893633, + 0.005049658007919788, + 0.007788886781781912, + 0.006624043453484774, + 0.006266366224735975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.00463874451816082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054945044219493866 + ], + "opposite_logprob_tokens": [ + -2.9299449920654297 + ], + "logprob_reward": -0.054945044219493866, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054945044219493866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1486756056547165, + 0.062491223216056824, + 0.1544969081878662, + 0.22447195649147034, + 0.05287303030490875, + 0.19085069000720978, + 0.03975927084684372, + 0.047212161123752594, + 0.028544185683131218, + 0.05062498152256012 + ], + "credit_map": [ + 0.017276683822274208, + 0.02015097439289093, + 0.023398609831929207, + 0.027346830815076828, + 0.018494514748454094, + 0.012866387143731117, + 0.014898248948156834, + 0.014243362471461296, + 0.013293171301484108, + 0.012095019221305847, + 0.006042064633220434, + 0.004737960174679756, + 0.007260903716087341, + 0.006274443119764328, + 0.005648895166814327, + 0.0071387686766684055, + 0.018048642203211784, + 0.024923253804445267, + 0.015628762543201447, + 0.014788588508963585, + 0.031688254326581955, + 0.01512632705271244, + 0.019857896491885185, + 0.014435186982154846, + 0.036751970648765564, + 0.03901848569512367, + 0.028962893411517143, + 0.019642366096377373, + 0.016050344333052635, + 0.03502003476023674, + 0.015393550507724285, + 0.033632297068834305, + 0.005621148273348808, + 0.0060414778999984264, + 0.005578465294092894, + 0.004721138160675764, + 0.006708188913762569, + 0.00849360041320324, + 0.007859054021537304, + 0.007849961519241333, + 0.023703087121248245, + 0.016374681144952774, + 0.0376649834215641, + 0.015542102046310902, + 0.02450357750058174, + 0.035880040377378464, + 0.01300114020705223, + 0.024181079119443893, + 0.0054397545754909515, + 0.008793199434876442, + 0.004286790732294321, + 0.0033531272783875465, + 0.006170495878905058, + 0.005280632991343737, + 0.0030277343466877937, + 0.00340753560885787, + 0.004022226668894291, + 0.005199408158659935, + 0.005513824988156557, + 0.006565216463059187, + 0.006434576120227575, + 0.005048383492976427, + 0.007025048602372408, + 0.007403474301099777, + 0.0024735729675740004, + 0.00231781299225986, + 0.003716734005138278, + 0.002462042262777686, + 0.006155744194984436, + 0.003195419441908598, + 0.0023670268710702658, + 0.00585583271458745, + 0.004898412618786097, + 0.006263016723096371, + 0.009135124273598194, + 0.006268922705203295, + 0.006903988774865866, + 0.006141239311546087, + 0.00562550313770771, + 0.0053887739777565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.054945044219493866 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage steam train with jet-black metal and red drive wheels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the train moving at high speed through a deep, precipitous valley flanked by steep slopes with lush dark green pine forests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are thick, billowing plumes of white steam erupting from the train's towering chimneys?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a golden morning sun piercing the mist, tinging the clouds with a warm golden edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a dynamic low-angle tracking shot following the train?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062015119940042496 + ], + "opposite_logprob_tokens": [ + -2.8120150566101074 + ], + "logprob_reward": -0.062015119940042496, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062015119940042496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10918544232845306, + 0.07589422911405563, + 0.1687658280134201, + 0.2606062889099121, + 0.034860897809267044, + 0.048688169568777084, + 0.060431402176618576, + 0.05314085632562637, + 0.13046640157699585, + 0.05796045809984207 + ], + "credit_map": [ + 0.016275526955723763, + 0.015388758853077888, + 0.01595156453549862, + 0.016468841582536697, + 0.013399675488471985, + 0.011831741780042648, + 0.010634648613631725, + 0.00923468079417944, + 0.017563918605446815, + 0.015016884543001652, + 0.006443039979785681, + 0.005499829538166523, + 0.005273624788969755, + 0.005111651495099068, + 0.010536274872720242, + 0.010449003428220749, + 0.017893420532345772, + 0.02738981507718563, + 0.02113722823560238, + 0.014230542816221714, + 0.023521067574620247, + 0.020861078053712845, + 0.01781046763062477, + 0.025922205299139023, + 0.03519998863339424, + 0.018390702083706856, + 0.013007896021008492, + 0.010903826914727688, + 0.017927827313542366, + 0.06140723451972008, + 0.02660805732011795, + 0.07716076821088791, + 0.005127201322466135, + 0.005097707733511925, + 0.0028674951754510403, + 0.002405549166724086, + 0.005681994371116161, + 0.006474432535469532, + 0.0032072567846626043, + 0.003999259322881699, + 0.004910092800855637, + 0.005456306505948305, + 0.008453520946204662, + 0.005161155480891466, + 0.007877829484641552, + 0.005621580872684717, + 0.005460307467728853, + 0.00574737461283803, + 0.0063772969879209995, + 0.0065306113101542, + 0.0044942256063222885, + 0.008098207414150238, + 0.010697470977902412, + 0.007881058380007744, + 0.00926215574145317, + 0.007090373896062374, + 0.005697950720787048, + 0.013824363239109516, + 0.004851134493947029, + 0.0044373092241585255, + 0.005022225435823202, + 0.007255639880895615, + 0.004408045671880245, + 0.007644190452992916, + 0.010500835254788399, + 0.009884022176265717, + 0.010057120583951473, + 0.01740344613790512, + 0.031018750742077827, + 0.012756971642374992, + 0.010257224552333355, + 0.028588037937879562, + 0.005545463878661394, + 0.0055976808071136475, + 0.009427250362932682, + 0.007780220825225115, + 0.006610256154090166, + 0.007899499498307705, + 0.0073471576906740665, + 0.0077529288828372955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.062015119940042496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_13/index_tensor([3856, 3856])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_13/index_tensor([3856, 3856])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5001355969dd6749f312763115e5a0f33650b110 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_13/index_tensor([3856, 3856])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a horizontal angle, the image exudes Rembrandt-like light and shadow, reminiscent of a classical oil painting, its brushstrokes intense and dense. The sole illumination comes from a flickering candlelight in the lower center of the frame. This thick, half-burned candle rests in a rustic clay candlestick, its warm orange glow the centerpiece of the dim scene. The light gently illuminates the face of an elderly woman, whose deep wrinkles, like a dry riverbed, crisscross across her forehead, eyes, and cheeks, each bearing a story. Her silver-white hair is loosely tied back with a simple wooden hairpin, with a few strands falling by her cheeks. She wears a dark brown, rough linen top. She leans slightly forward, her cloudy yet luminous eyes fixed intently on the dancing flame, her lips moving at an almost inaudible rhythm, as if engaged in a private conversation with time or a deity. The candlelight casts vivid shadows on her deep wrinkles, trembling slightly with the flickering flame, while the background is completely swallowed by the deep and indissoluble darkness, creating an atmosphere of tranquility, solemnity, contemplation and mystery.", + "seed": 26866, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_13/index_3856_seed_26866_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03351480755954981, + "vlm_logprob": -0.03351480755954981, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03351480755954981, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006359340623021126 + ], + "opposite_logprob_tokens": [ + -5.381359577178955 + ], + "logprob_reward": -0.006359340623021126, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006359340623021126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2151130586862564, + 0.1200617104768753, + 0.05154058337211609, + 0.05209805443882942, + 0.17012068629264832, + 0.1467067152261734, + 0.033960603177547455, + 0.03781219944357872, + 0.05472993105649948, + 0.11785642802715302 + ], + "credit_map": [ + 0.023222194984555244, + 0.017440544441342354, + 0.04894353076815605, + 0.016415998339653015, + 0.027760546654462814, + 0.03661040961742401, + 0.022768579423427582, + 0.021951254457235336, + 0.011270382441580296, + 0.014558844268321991, + 0.017787009477615356, + 0.012158687226474285, + 0.019614169374108315, + 0.016593221575021744, + 0.015532604418694973, + 0.0125467823818326, + 0.008881552144885063, + 0.007186487782746553, + 0.007389658130705357, + 0.003935201093554497, + 0.0048720864579081535, + 0.008342315442860126, + 0.004444564692676067, + 0.006488717161118984, + 0.004128208849579096, + 0.0072260391898453236, + 0.008164986036717892, + 0.004461678676307201, + 0.006475473288446665, + 0.0093491580337286, + 0.006147809326648712, + 0.0061447033658623695, + 0.018212633207440376, + 0.028252482414245605, + 0.02123847045004368, + 0.02192148193717003, + 0.02290799655020237, + 0.022665169090032578, + 0.017115311697125435, + 0.017807144671678543, + 0.030871056020259857, + 0.024900007992982864, + 0.020569831132888794, + 0.013811908662319183, + 0.012880680151283741, + 0.01709982566535473, + 0.01112404651939869, + 0.015449371188879013, + 0.0052200122736394405, + 0.00424770126119256, + 0.00446440139785409, + 0.0028295426163822412, + 0.003976595588028431, + 0.00576374726369977, + 0.003574956441298127, + 0.003883649129420519, + 0.004498342517763376, + 0.005813567899167538, + 0.004729702137410641, + 0.003499878104776144, + 0.006599533371627331, + 0.004839703440666199, + 0.003945204429328442, + 0.0038862701039761305, + 0.007214508485049009, + 0.011289305053651333, + 0.009386316873133183, + 0.0063329641707241535, + 0.004911789204925299, + 0.006594898644834757, + 0.003503707004711032, + 0.005496439058333635, + 0.008548739366233349, + 0.01868690736591816, + 0.022418079897761345, + 0.014215920120477676, + 0.012968337163329124, + 0.018028030171990395, + 0.008793608285486698, + 0.014196805655956268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.006359340623021126 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649304784834385 + ], + "opposite_logprob_tokens": [ + -5.3796491622924805 + ], + "logprob_reward": -0.004649304784834385, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649304784834385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1679108738899231, + 0.1046755239367485, + 0.08049143850803375, + 0.07430551201105118, + 0.16153448820114136, + 0.08120258152484894, + 0.03418191894888878, + 0.058857474476099014, + 0.023740636184811592, + 0.21309953927993774 + ], + "credit_map": [ + 0.015383693389594555, + 0.024890916422009468, + 0.025265099480748177, + 0.01729983277618885, + 0.024918271228671074, + 0.026479005813598633, + 0.01847533881664276, + 0.01519870851188898, + 0.008282042108476162, + 0.015167133882641792, + 0.01754913479089737, + 0.010112752206623554, + 0.013617834076285362, + 0.016131987795233727, + 0.012662471272051334, + 0.011152166873216629, + 0.006857702508568764, + 0.009150002151727676, + 0.01291366945952177, + 0.00869423896074295, + 0.009693048894405365, + 0.01425123866647482, + 0.009902499616146088, + 0.009029028937220573, + 0.005945703946053982, + 0.010979292914271355, + 0.01362281758338213, + 0.007103014271706343, + 0.009552424773573875, + 0.010519464500248432, + 0.007635734509676695, + 0.008947065100073814, + 0.015096264891326427, + 0.024216771125793457, + 0.024618297815322876, + 0.021764736622571945, + 0.021323613822460175, + 0.02053740806877613, + 0.018113596364855766, + 0.015863800421357155, + 0.013314522802829742, + 0.00932731106877327, + 0.013985813595354557, + 0.00708915526047349, + 0.0072615197859704494, + 0.012124772183597088, + 0.007799102459102869, + 0.010300385765731335, + 0.005645119585096836, + 0.004393056966364384, + 0.004772295709699392, + 0.0026948570739477873, + 0.0034061248879879713, + 0.005567338317632675, + 0.003999626729637384, + 0.0037035003770142794, + 0.00690300902351737, + 0.008778313174843788, + 0.007653259206563234, + 0.006546275224536657, + 0.009931975044310093, + 0.007602063938975334, + 0.0059214867651462555, + 0.00552109070122242, + 0.0032362372148782015, + 0.00348491920158267, + 0.0034429472871124744, + 0.002684100763872266, + 0.002249601297080517, + 0.003185752546414733, + 0.003229591529816389, + 0.002227486576884985, + 0.015834402292966843, + 0.02670007385313511, + 0.03269157558679581, + 0.02083299122750759, + 0.021572083234786987, + 0.036158833652734756, + 0.024698276072740555, + 0.034611307084560394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.004649304784834385 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497235059738159 + ], + "opposite_logprob_tokens": [ + -2.9299724102020264 + ], + "logprob_reward": -0.05497235059738159, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497235059738159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2731809616088867, + 0.09165102988481522, + 0.09936539828777313, + 0.06735720485448837, + 0.10559738427400589, + 0.05940889194607735, + 0.06447022408246994, + 0.04826509207487106, + 0.060652315616607666, + 0.13005146384239197 + ], + "credit_map": [ + 0.03762127086520195, + 0.031172750517725945, + 0.053775347769260406, + 0.02156873792409897, + 0.047180596739053726, + 0.036516692489385605, + 0.02130579948425293, + 0.02403976395726204, + 0.01025623269379139, + 0.016330275684595108, + 0.012444533407688141, + 0.00864369235932827, + 0.012343281880021095, + 0.014160992577672005, + 0.006595057435333729, + 0.010876966640353203, + 0.012357130646705627, + 0.012062418274581432, + 0.021614646539092064, + 0.008981496095657349, + 0.011125097051262856, + 0.0158119834959507, + 0.006873604375869036, + 0.010539018549025059, + 0.008579956367611885, + 0.009401539340615273, + 0.012832843698561192, + 0.006883018650114536, + 0.007184648886322975, + 0.010727503336966038, + 0.00545109249651432, + 0.006296598352491856, + 0.00949176773428917, + 0.018754836171865463, + 0.015624087303876877, + 0.012605438940227032, + 0.012870844453573227, + 0.013117466121912003, + 0.011218942701816559, + 0.011913998983800411, + 0.01058737188577652, + 0.006458053831011057, + 0.010415453463792801, + 0.005718010012060404, + 0.003965693060308695, + 0.008344746194779873, + 0.005628403276205063, + 0.008291158825159073, + 0.007600659038871527, + 0.0095302639529109, + 0.009364261291921139, + 0.0070073725655674934, + 0.007167010568082333, + 0.010693519376218319, + 0.005318991374224424, + 0.007788143586367369, + 0.005761292763054371, + 0.007359338458627462, + 0.0076623097993433475, + 0.005584476515650749, + 0.006630371790379286, + 0.005710477475076914, + 0.004842320457100868, + 0.004714503418654203, + 0.006602542474865913, + 0.008940484374761581, + 0.010093615390360355, + 0.007864419370889664, + 0.006885986775159836, + 0.007563647348433733, + 0.0056072622537612915, + 0.007094359956681728, + 0.011509966105222702, + 0.01894986443221569, + 0.019573289901018143, + 0.018354782834649086, + 0.01573859713971615, + 0.01727239042520523, + 0.012128408998250961, + 0.016524160280823708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.05497235059738159 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538531795144081 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538531795144081, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538531795144081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32251498103141785, + 0.058159828186035156, + 0.05235578119754791, + 0.033880565315485, + 0.12417026609182358, + 0.23957571387290955, + 0.04555986821651459, + 0.025282133370637894, + 0.029190270230174065, + 0.06931062042713165 + ], + "credit_map": [ + 0.03254551440477371, + 0.04050245136022568, + 0.055940888822078705, + 0.023904649540781975, + 0.06262949109077454, + 0.044630203396081924, + 0.03335846588015556, + 0.029003312811255455, + 0.005654162727296352, + 0.007804054766893387, + 0.01106164138764143, + 0.006682905368506908, + 0.006667847745120525, + 0.00808512233197689, + 0.0059332652017474174, + 0.006270833313465118, + 0.004167248494923115, + 0.008286423981189728, + 0.010285751894116402, + 0.00445558363571763, + 0.005964487791061401, + 0.007743629161268473, + 0.004888925235718489, + 0.006563729140907526, + 0.00394164165481925, + 0.003995027858763933, + 0.007281774189323187, + 0.002918278332799673, + 0.0034461033064872026, + 0.005394428968429565, + 0.0030230439733713865, + 0.0038802679628133774, + 0.012467484921216965, + 0.01870656944811344, + 0.01937211863696575, + 0.014831655658781528, + 0.0156724750995636, + 0.016188576817512512, + 0.011432195082306862, + 0.015499192290008068, + 0.038019854575395584, + 0.033218178898096085, + 0.041024964302778244, + 0.02419724501669407, + 0.025912417098879814, + 0.03418919816613197, + 0.016641540452837944, + 0.026372311636805534, + 0.006296704988926649, + 0.0067053441889584064, + 0.006174534093588591, + 0.004296410828828812, + 0.004900938831269741, + 0.00745247071608901, + 0.004261732567101717, + 0.005471731070429087, + 0.003024389734491706, + 0.003379038069397211, + 0.0031080907210707664, + 0.002264169277623296, + 0.004946279339492321, + 0.003349235514178872, + 0.002695741830393672, + 0.0025151895824819803, + 0.004126356914639473, + 0.005240078084170818, + 0.004576835781335831, + 0.003291364060714841, + 0.002698390046134591, + 0.0037948533426970243, + 0.0019587953574955463, + 0.0035035978071391582, + 0.005809547379612923, + 0.009601698257029057, + 0.012589260935783386, + 0.007209789473563433, + 0.006648653652518988, + 0.009929509833455086, + 0.0075663053430616856, + 0.009955854155123234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.012538531795144081 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905450999736786 + ], + "opposite_logprob_tokens": [ + -2.464054584503174 + ], + "logprob_reward": -0.08905450999736786, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905450999736786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23146091401576996, + 0.17154192924499512, + 0.08026125282049179, + 0.11037428677082062, + 0.10459389537572861, + 0.09474322199821472, + 0.05485685169696808, + 0.04622812941670418, + 0.01408217754215002, + 0.09185724705457687 + ], + "credit_map": [ + 0.019825812429189682, + 0.030872246250510216, + 0.05756361782550812, + 0.020805252715945244, + 0.01816965639591217, + 0.04111848771572113, + 0.017477484419941902, + 0.025628378614783287, + 0.015213553793728352, + 0.018531451001763344, + 0.027941890060901642, + 0.015726732090115547, + 0.02586541324853897, + 0.026546843349933624, + 0.022448277100920677, + 0.01926778256893158, + 0.011406484991312027, + 0.008800930343568325, + 0.015555443242192268, + 0.007534907665103674, + 0.007292316760867834, + 0.011829163879156113, + 0.007384457625448704, + 0.010457556694746017, + 0.015208298340439796, + 0.015571247786283493, + 0.015027196146547794, + 0.00822279043495655, + 0.0158168263733387, + 0.019621137529611588, + 0.009335200302302837, + 0.01157160010188818, + 0.015498452819883823, + 0.014405250549316406, + 0.02004298008978367, + 0.00825295690447092, + 0.011126851662993431, + 0.014200649224221706, + 0.008448245003819466, + 0.01261850818991661, + 0.01651511713862419, + 0.012756388634443283, + 0.014855997636914253, + 0.009536180645227432, + 0.010582971386611462, + 0.01328672468662262, + 0.007307811640202999, + 0.009902036748826504, + 0.009074412286281586, + 0.007742151152342558, + 0.008284283801913261, + 0.004579451400786638, + 0.005879533011466265, + 0.008601317182183266, + 0.004691043868660927, + 0.006004660855978727, + 0.006464431528002024, + 0.00689836498349905, + 0.007014071568846703, + 0.004755240399390459, + 0.006274424027651548, + 0.005310334265232086, + 0.004879101645201445, + 0.004632165189832449, + 0.0019210922764614224, + 0.0015371269546449184, + 0.002216567052528262, + 0.001194790005683899, + 0.001186492620036006, + 0.00256960466504097, + 0.0015216863248497248, + 0.0019348189234733582, + 0.007742017041891813, + 0.013472933322191238, + 0.011831071227788925, + 0.01083445642143488, + 0.010367264039814472, + 0.014244618825614452, + 0.010145401582121849, + 0.013219497166574001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.08905450999736786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a horizontal angle, the image exudes Rembrandt-like light and shadow, reminiscent of a classical oil painting, its brushstrokes intense and dense. The sole illumination comes from a flickering candlelight in the lower center of the frame. This thick, half-burned candle rests in a rustic clay candlestick, its warm orange glow the centerpiece of the dim scene. The light gently illuminates the face of an elderly woman, whose deep wrinkles, like a dry riverbed, crisscross across her forehead, eyes, and cheeks, each bearing a story. Her silver-white hair is loosely tied back with a simple wooden hairpin, with a few strands falling by her cheeks. She wears a dark brown, rough linen top. She leans slightly forward, her cloudy yet luminous eyes fixed intently on the dancing flame, her lips moving at an almost inaudible rhythm, as if engaged in a private conversation with time or a deity. The candlelight casts vivid shadows on her deep wrinkles, trembling slightly with the flickering flame, while the background is completely swallowed by the deep and indissoluble darkness, creating an atmosphere of tranquility, solemnity, contemplation and mystery.", + "seed": 26867, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_13/index_3856_seed_26867_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04502465296536684, + "vlm_logprob": -0.04502465296536684, + "vlm_yes_no_margin": 3.774999976158142, + "vlm_token_credit": -0.04502465296536684, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005700284615159035 + ], + "opposite_logprob_tokens": [ + -5.75570011138916 + ], + "logprob_reward": -0.005700284615159035, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.005700284615159035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32488390803337097, + 0.1174139603972435, + 0.10637900233268738, + 0.08445605635643005, + 0.06086226552724838, + 0.0586659200489521, + 0.06614688038825989, + 0.06413879990577698, + 0.05278942734003067, + 0.06426376849412918 + ], + "credit_map": [ + 0.02170652523636818, + 0.04823101684451103, + 0.03484519571065903, + 0.03980989754199982, + 0.0649348571896553, + 0.042126137763261795, + 0.05017033591866493, + 0.023059943690896034, + 0.010008908808231354, + 0.01450891513377428, + 0.017897658050060272, + 0.019831575453281403, + 0.014098197221755981, + 0.016241153702139854, + 0.016227293759584427, + 0.008600256405770779, + 0.009756321087479591, + 0.01789049431681633, + 0.01605279929935932, + 0.017848212271928787, + 0.01355009339749813, + 0.015181216411292553, + 0.01053051371127367, + 0.005569350905716419, + 0.005883514415472746, + 0.016657499596476555, + 0.009251452051103115, + 0.011226439848542213, + 0.015460182912647724, + 0.010812176391482353, + 0.00894459430128336, + 0.0062202028930187225, + 0.00527859665453434, + 0.007062130607664585, + 0.011233780533075333, + 0.008217589929699898, + 0.007451124954968691, + 0.011336766183376312, + 0.006583517882972956, + 0.003698764368891716, + 0.004262098576873541, + 0.008369401097297668, + 0.00890097115188837, + 0.00822535902261734, + 0.007141738664358854, + 0.007938704453408718, + 0.0084496745839715, + 0.005377973895519972, + 0.00462749321013689, + 0.008923106826841831, + 0.010702395811676979, + 0.009315362200140953, + 0.008505508303642273, + 0.009325646795332432, + 0.010155457071959972, + 0.004591910634189844, + 0.005727281793951988, + 0.007766504306346178, + 0.010345197282731533, + 0.00896050501614809, + 0.00683281384408474, + 0.009307321161031723, + 0.00893794372677803, + 0.006261235103011131, + 0.003781303996220231, + 0.007384815253317356, + 0.006270694546401501, + 0.0068128216080367565, + 0.005116297863423824, + 0.007694989908486605, + 0.007530424278229475, + 0.008198081515729427, + 0.007342995144426823, + 0.00824167300015688, + 0.012327936477959156, + 0.00794979464262724, + 0.007721943315118551, + 0.00918408203870058, + 0.0077301147393882275, + 0.003765231231227517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005700284615159035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016092415899038315 + ], + "opposite_logprob_tokens": [ + -4.141092300415039 + ], + "logprob_reward": -0.016092415899038315, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016092415899038315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20505870878696442, + 0.11145862191915512, + 0.09471221268177032, + 0.11709406226873398, + 0.08251716941595078, + 0.11584936827421188, + 0.042844295501708984, + 0.082460418343544, + 0.057825859636068344, + 0.09017924219369888 + ], + "credit_map": [ + 0.020921912044286728, + 0.02576015144586563, + 0.032233450561761856, + 0.03511935472488403, + 0.024143656715750694, + 0.027203138917684555, + 0.023198530077934265, + 0.016478506848216057, + 0.010001691989600658, + 0.01382698118686676, + 0.01766636036336422, + 0.01407702174037695, + 0.012304007075726986, + 0.016280807554721832, + 0.014248259365558624, + 0.01305349636822939, + 0.007974731735885143, + 0.010030427947640419, + 0.013239376246929169, + 0.015321241691708565, + 0.012538252398371696, + 0.013285182416439056, + 0.012415368109941483, + 0.009907633997499943, + 0.00910054799169302, + 0.01396025437861681, + 0.016569403931498528, + 0.013153433799743652, + 0.012086895294487476, + 0.01633121632039547, + 0.015559036284685135, + 0.02033327706158161, + 0.006702878046780825, + 0.010471773333847523, + 0.011976182460784912, + 0.010584468021988869, + 0.012131703086197376, + 0.011075113900005817, + 0.008923095650970936, + 0.010651953518390656, + 0.009798352606594563, + 0.013889815658330917, + 0.017359038814902306, + 0.01446093712002039, + 0.016210826113820076, + 0.016227316111326218, + 0.012810301035642624, + 0.01509278453886509, + 0.0027850775513798, + 0.004534456413239241, + 0.00557099562138319, + 0.005584297701716423, + 0.005798707716166973, + 0.0062074679881334305, + 0.0070404717698693275, + 0.0053228228352963924, + 0.005773270037025213, + 0.008228168822824955, + 0.013683531433343887, + 0.009286933578550816, + 0.006534980144351721, + 0.012476596049964428, + 0.015982966870069504, + 0.010493974201381207, + 0.004704191815108061, + 0.005662727169692516, + 0.008211921900510788, + 0.007243957370519638, + 0.005509380716830492, + 0.00749611621722579, + 0.00913692731410265, + 0.009860635735094547, + 0.008122687228024006, + 0.011059291660785675, + 0.01378649938851595, + 0.01268398854881525, + 0.01207926869392395, + 0.011101660318672657, + 0.012917391955852509, + 0.008428463712334633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.016092415899038315 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206251308321953 + ], + "opposite_logprob_tokens": [ + -2.8120625019073486 + ], + "logprob_reward": -0.06206251308321953, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206251308321953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30816876888275146, + 0.14417724311351776, + 0.12119462341070175, + 0.05412694811820984, + 0.05727500095963478, + 0.06442590802907944, + 0.05986112728714943, + 0.08910629153251648, + 0.05193101242184639, + 0.049733057618141174 + ], + "credit_map": [ + 0.0301357414573431, + 0.05366963893175125, + 0.04262767359614372, + 0.04114138334989548, + 0.04206405207514763, + 0.03826272115111351, + 0.03693949431180954, + 0.023328062146902084, + 0.014771759510040283, + 0.019581792876124382, + 0.02102532796561718, + 0.024094855412840843, + 0.01603640802204609, + 0.01959250308573246, + 0.01599034108221531, + 0.013084250502288342, + 0.009598733857274055, + 0.016148725524544716, + 0.017289038747549057, + 0.022733286023139954, + 0.015239098109304905, + 0.01664132811129093, + 0.013517182320356369, + 0.010027235373854637, + 0.005433486308902502, + 0.004999429918825626, + 0.007048758678138256, + 0.005638208705931902, + 0.006731491535902023, + 0.008474119007587433, + 0.006660000886768103, + 0.009141453541815281, + 0.004768808372318745, + 0.004844179842621088, + 0.011325070634484291, + 0.006010370794683695, + 0.005786076653748751, + 0.009394939988851547, + 0.008072342723608017, + 0.007073213811963797, + 0.005360679235309362, + 0.007718848064541817, + 0.009730597026646137, + 0.008972026407718658, + 0.009882532060146332, + 0.00865093618631363, + 0.009744894690811634, + 0.004365391097962856, + 0.004455962218344212, + 0.007835234515368938, + 0.009709387086331844, + 0.008478393778204918, + 0.008250124752521515, + 0.008572518825531006, + 0.009221644140779972, + 0.003337860805913806, + 0.008680342696607113, + 0.012235740199685097, + 0.014348119497299194, + 0.013318073935806751, + 0.010190282948315144, + 0.013253013603389263, + 0.011580243706703186, + 0.005500473082065582, + 0.004110127687454224, + 0.005294526927173138, + 0.008201668038964272, + 0.006859800778329372, + 0.00534099293872714, + 0.007682927884161472, + 0.007498160935938358, + 0.006942806299775839, + 0.0037572323344647884, + 0.008930007927119732, + 0.006375318858772516, + 0.008667073212563992, + 0.005857846699655056, + 0.005575647111982107, + 0.006060061510652304, + 0.00450986810028553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.06206251308321953 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014198473654687405 + ], + "opposite_logprob_tokens": [ + -4.264198303222656 + ], + "logprob_reward": -0.014198473654687405, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014198473654687405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29312247037887573, + 0.07623982429504395, + 0.05557560548186302, + 0.044506870210170746, + 0.10177778452634811, + 0.05643710866570473, + 0.049767374992370605, + 0.18441079556941986, + 0.0537029504776001, + 0.08445920795202255 + ], + "credit_map": [ + 0.03708554431796074, + 0.04009168595075607, + 0.04539651423692703, + 0.03169833496212959, + 0.03333484008908272, + 0.04151654988527298, + 0.038738708943128586, + 0.02526029571890831, + 0.006316969636827707, + 0.008290481753647327, + 0.009918254800140858, + 0.012533444911241531, + 0.007409217301756144, + 0.010791289620101452, + 0.009880228899419308, + 0.011099940165877342, + 0.005393635015934706, + 0.007672784850001335, + 0.008698510937392712, + 0.008281280286610126, + 0.006589445751160383, + 0.007778341416269541, + 0.007531789597123861, + 0.0036298162303864956, + 0.0034257674124091864, + 0.004853367805480957, + 0.007536816876381636, + 0.005985659547150135, + 0.006576883140951395, + 0.006075198296457529, + 0.005594135727733374, + 0.004459040239453316, + 0.008622691035270691, + 0.010483531281352043, + 0.01412515714764595, + 0.011681714095175266, + 0.016435125842690468, + 0.015545511618256569, + 0.01425970159471035, + 0.010624351911246777, + 0.005117825232446194, + 0.007943126372992992, + 0.009916889481246471, + 0.007909199222922325, + 0.006834258791059256, + 0.007472184021025896, + 0.007916996255517006, + 0.003326622536405921, + 0.0034622836392372847, + 0.00666787987574935, + 0.006588580086827278, + 0.007764066569507122, + 0.008357190527021885, + 0.006436111871153116, + 0.005968788173049688, + 0.004522474482655525, + 0.014787289313971996, + 0.02582463063299656, + 0.029241494834423065, + 0.027023550122976303, + 0.02049129456281662, + 0.027091369032859802, + 0.030141735449433327, + 0.00980942789465189, + 0.004651904571801424, + 0.006906681694090366, + 0.007743120659142733, + 0.00659699272364378, + 0.006089433096349239, + 0.00736815482378006, + 0.00717131095007062, + 0.007175352890044451, + 0.011509883217513561, + 0.011804257519543171, + 0.013073011301457882, + 0.011569423601031303, + 0.009925372898578644, + 0.010172858834266663, + 0.010812904685735703, + 0.005591495428234339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.014198473654687405 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706957757472992 + ], + "opposite_logprob_tokens": [ + -2.1270694732666016 + ], + "logprob_reward": -0.12706957757472992, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12706957757472992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25059711933135986, + 0.07873346656560898, + 0.1342022716999054, + 0.09631020575761795, + 0.12945441901683807, + 0.06238742545247078, + 0.04667118564248085, + 0.09235440194606781, + 0.06364329159259796, + 0.04564623162150383 + ], + "credit_map": [ + 0.018961336463689804, + 0.03412369266152382, + 0.037431519478559494, + 0.03520532324910164, + 0.03545605391263962, + 0.035800497978925705, + 0.03337525948882103, + 0.020243434235453606, + 0.007123738527297974, + 0.007452929858118296, + 0.013392360880970955, + 0.013114742003381252, + 0.00788105372339487, + 0.011927099898457527, + 0.009820844046771526, + 0.008020691573619843, + 0.010210319422185421, + 0.018010545521974564, + 0.019185753539204597, + 0.02209601178765297, + 0.0191543810069561, + 0.01836232654750347, + 0.018407385796308517, + 0.008775551803410053, + 0.008304561488330364, + 0.013877375982701778, + 0.01280276756733656, + 0.015277869068086147, + 0.013549664057791233, + 0.012951352633535862, + 0.012190509587526321, + 0.007356107700616121, + 0.010483291000127792, + 0.013085260055959225, + 0.020582983270287514, + 0.019381463527679443, + 0.01475248858332634, + 0.019741179421544075, + 0.019488558173179626, + 0.011939186602830887, + 0.004739797208458185, + 0.009485900402069092, + 0.008417271077632904, + 0.011076480150222778, + 0.008807715028524399, + 0.007787775248289108, + 0.00776392687112093, + 0.0043085613287985325, + 0.003805631073191762, + 0.003961564507335424, + 0.006556812673807144, + 0.008636794984340668, + 0.0055402531288564205, + 0.006944681517779827, + 0.006146001163870096, + 0.005079448223114014, + 0.005877514835447073, + 0.015231875702738762, + 0.014488817192614079, + 0.01441203523427248, + 0.0103689543902874, + 0.01304525975137949, + 0.012940586544573307, + 0.005989359226077795, + 0.005291988607496023, + 0.00666632829234004, + 0.010776882991194725, + 0.008899254724383354, + 0.005265546031296253, + 0.00981945637613535, + 0.009075067937374115, + 0.007848766632378101, + 0.005010678432881832, + 0.0053606485016644, + 0.007840743288397789, + 0.007132760714739561, + 0.004293512552976608, + 0.00654488755390048, + 0.006100176367908716, + 0.003362820250913501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.12706957757472992 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a horizontal angle, the image exudes Rembrandt-like light and shadow, reminiscent of a classical oil painting, its brushstrokes intense and dense. The sole illumination comes from a flickering candlelight in the lower center of the frame. This thick, half-burned candle rests in a rustic clay candlestick, its warm orange glow the centerpiece of the dim scene. The light gently illuminates the face of an elderly woman, whose deep wrinkles, like a dry riverbed, crisscross across her forehead, eyes, and cheeks, each bearing a story. Her silver-white hair is loosely tied back with a simple wooden hairpin, with a few strands falling by her cheeks. She wears a dark brown, rough linen top. She leans slightly forward, her cloudy yet luminous eyes fixed intently on the dancing flame, her lips moving at an almost inaudible rhythm, as if engaged in a private conversation with time or a deity. The candlelight casts vivid shadows on her deep wrinkles, trembling slightly with the flickering flame, while the background is completely swallowed by the deep and indissoluble darkness, creating an atmosphere of tranquility, solemnity, contemplation and mystery.", + "seed": 26868, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_13/index_3856_seed_26868_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.044848608784377575, + "vlm_logprob": -0.044848608784377575, + "vlm_yes_no_margin": 3.924999976158142, + "vlm_token_credit": -0.044848608784377575, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006305443122982979 + ], + "opposite_logprob_tokens": [ + -5.50630521774292 + ], + "logprob_reward": -0.006305443122982979, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006305443122982979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21816565096378326, + 0.26003026962280273, + 0.0531981885433197, + 0.10110899806022644, + 0.0818304643034935, + 0.1402810513973236, + 0.058644313365221024, + 0.03862517327070236, + 0.021666985005140305, + 0.026448991149663925 + ], + "credit_map": [ + 0.022040899842977524, + 0.0401747040450573, + 0.026607710868120193, + 0.03832981362938881, + 0.03439760580658913, + 0.024895016103982925, + 0.020268766209483147, + 0.011451133526861668, + 0.024502350017428398, + 0.0423206202685833, + 0.030646616593003273, + 0.05558903142809868, + 0.03598930314183235, + 0.037461064755916595, + 0.022646969184279442, + 0.010874295607209206, + 0.004663880914449692, + 0.011523925699293613, + 0.007859181612730026, + 0.01068165060132742, + 0.0053912317380309105, + 0.0061243800446391106, + 0.004185752477496862, + 0.0027681842911988497, + 0.009401120245456696, + 0.015201829373836517, + 0.012699289247393608, + 0.016784869134426117, + 0.01874019391834736, + 0.01254274882376194, + 0.010746192187070847, + 0.004992758389562368, + 0.006935506127774715, + 0.015329443849623203, + 0.010087858885526657, + 0.015532514080405235, + 0.010646514594554901, + 0.010000106878578663, + 0.007904937490820885, + 0.005393580999225378, + 0.012591679580509663, + 0.02477497234940529, + 0.016261447221040726, + 0.032042067497968674, + 0.016687147319316864, + 0.01688833348453045, + 0.013135016895830631, + 0.007900395430624485, + 0.004062870983034372, + 0.01013385970145464, + 0.010946149006485939, + 0.011264936067163944, + 0.006190930027514696, + 0.007475812919437885, + 0.004554972518235445, + 0.004014780279248953, + 0.0037474508862942457, + 0.005459485575556755, + 0.006737787742167711, + 0.007578345015645027, + 0.0054872590117156506, + 0.005306775216013193, + 0.002727376762777567, + 0.0015806917799636722, + 0.0018505098996683955, + 0.0032946355640888214, + 0.002877091523259878, + 0.003836284391582012, + 0.003003291552886367, + 0.0032958851661533117, + 0.0017635964322835207, + 0.0017456916393712163, + 0.002717679599300027, + 0.004512356594204903, + 0.0035160756669938564, + 0.005465705879032612, + 0.0033941559959203005, + 0.0035936767235398293, + 0.0017127158353105187, + 0.0015366260195150971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.006305443122982979 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005965879186987877 + ], + "opposite_logprob_tokens": [ + -5.130965709686279 + ], + "logprob_reward": -0.005965879186987877, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005965879186987877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11819300055503845, + 0.12991586327552795, + 0.18344593048095703, + 0.11898915469646454, + 0.06062644347548485, + 0.19410397112369537, + 0.04035625606775284, + 0.05859113112092018, + 0.0401347354054451, + 0.05564354360103607 + ], + "credit_map": [ + 0.010867479257285595, + 0.025898048654198647, + 0.014868086203932762, + 0.015404331497848034, + 0.013654333539307117, + 0.016187122091650963, + 0.010870159603655338, + 0.010443446226418018, + 0.014649251475930214, + 0.01942744106054306, + 0.015083652921020985, + 0.022755511105060577, + 0.020925244316458702, + 0.01822209358215332, + 0.011259107850492, + 0.007593561429530382, + 0.015476387925446033, + 0.02690916508436203, + 0.01926722750067711, + 0.03694494813680649, + 0.027797097340226173, + 0.025013552978634834, + 0.018633496016263962, + 0.013404077850282192, + 0.008934193290770054, + 0.014283210039138794, + 0.01525475550442934, + 0.025599248707294464, + 0.01922421343624592, + 0.014982543885707855, + 0.014333339408040047, + 0.006377649959176779, + 0.004595852456986904, + 0.007598630152642727, + 0.010334931313991547, + 0.009610866196453571, + 0.006967473309487104, + 0.008699620142579079, + 0.00790979340672493, + 0.00490927416831255, + 0.01846250332891941, + 0.026007844135165215, + 0.02667970024049282, + 0.03594371676445007, + 0.022875357419252396, + 0.03618316724896431, + 0.01369484979659319, + 0.014256836846470833, + 0.003736071987077594, + 0.005857165902853012, + 0.005440677050501108, + 0.00618591345846653, + 0.0060532065108418465, + 0.005672665778547525, + 0.0034505904186517, + 0.003959965892136097, + 0.004652019590139389, + 0.008084332570433617, + 0.007658491376787424, + 0.009025158360600471, + 0.0057289451360702515, + 0.008373728953301907, + 0.006307627540081739, + 0.008760825730860233, + 0.004362656269222498, + 0.004959873389452696, + 0.006328826304525137, + 0.00539898406714201, + 0.0058821444399654865, + 0.0056138308718800545, + 0.0037622174713760614, + 0.0038262002635747194, + 0.004905443172901869, + 0.008061373606324196, + 0.007732859812676907, + 0.011575622484087944, + 0.008653214201331139, + 0.007719549350440502, + 0.0038843918591737747, + 0.003111089812591672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.005965879186987877 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496230721473694 + ], + "opposite_logprob_tokens": [ + -2.929962396621704 + ], + "logprob_reward": -0.05496230721473694, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496230721473694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16011513769626617, + 0.1818966269493103, + 0.12878940999507904, + 0.14817745983600616, + 0.05028104782104492, + 0.07459892332553864, + 0.06466902792453766, + 0.1392924040555954, + 0.030505754053592682, + 0.021674150601029396 + ], + "credit_map": [ + 0.013685359619557858, + 0.029995612800121307, + 0.020807838067412376, + 0.03206741809844971, + 0.017855962738394737, + 0.021480901166796684, + 0.014420936815440655, + 0.009801113046705723, + 0.01190410740673542, + 0.03851642087101936, + 0.022780416533350945, + 0.03963587433099747, + 0.018910247832536697, + 0.028016218915581703, + 0.013400798663496971, + 0.008732542395591736, + 0.009501777589321136, + 0.025345399975776672, + 0.022894559428095818, + 0.02173701487481594, + 0.014798399992287159, + 0.01501611527055502, + 0.01187497191131115, + 0.007621187716722488, + 0.009474464692175388, + 0.037402063608169556, + 0.026349134743213654, + 0.027683543041348457, + 0.013724147342145443, + 0.01735197752714157, + 0.008290251716971397, + 0.007901892066001892, + 0.005058633163571358, + 0.006415719632059336, + 0.009115638211369514, + 0.00936941895633936, + 0.005009674467146397, + 0.006879427935928106, + 0.005113304127007723, + 0.00331923202611506, + 0.00816661398857832, + 0.00898116547614336, + 0.010189453139901161, + 0.016352428123354912, + 0.007629393599927425, + 0.013021724298596382, + 0.005948950070887804, + 0.004309201613068581, + 0.005011631175875664, + 0.015292948111891747, + 0.00691525312140584, + 0.010967566631734371, + 0.00969693623483181, + 0.008017891086637974, + 0.005802982021123171, + 0.00296382000669837, + 0.009834998287260532, + 0.040301986038684845, + 0.018227435648441315, + 0.02294430322945118, + 0.01612148806452751, + 0.020350204780697823, + 0.005650482606142759, + 0.005861521698534489, + 0.00237883604131639, + 0.0051934137009084225, + 0.0038444462697952986, + 0.0057671708054840565, + 0.003787535708397627, + 0.004531915299594402, + 0.0024134498089551926, + 0.0025889885146170855, + 0.002565732691437006, + 0.004709524568170309, + 0.0026972955092787743, + 0.0032075040508061647, + 0.0029417797923088074, + 0.0026556821539998055, + 0.0016056406311690807, + 0.0012909922515973449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.05496230721473694 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196357689797878 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196357689797878, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196357689797878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1727067232131958, + 0.11578623950481415, + 0.12997354567050934, + 0.09277253597974777, + 0.09443647414445877, + 0.21054776012897491, + 0.06278222054243088, + 0.047629307955503464, + 0.03917671740055084, + 0.03418848291039467 + ], + "credit_map": [ + 0.015889890491962433, + 0.02889764867722988, + 0.026810582727193832, + 0.03549192473292351, + 0.019798243418335915, + 0.021882442757487297, + 0.013204334303736687, + 0.010731653310358524, + 0.008119618520140648, + 0.023805854842066765, + 0.013273417949676514, + 0.020072704181075096, + 0.017895760014653206, + 0.015392765402793884, + 0.009864832274615765, + 0.007361291442066431, + 0.012842266820371151, + 0.019882261753082275, + 0.014948835596442223, + 0.02927720546722412, + 0.017132136970758438, + 0.016410907730460167, + 0.011773952282965183, + 0.007705969270318747, + 0.007013026624917984, + 0.021167917177081108, + 0.017559099942445755, + 0.015615873970091343, + 0.00929163582623005, + 0.010802664794027805, + 0.0060379719361662865, + 0.00528434244915843, + 0.008935182355344296, + 0.018430238589644432, + 0.014933038502931595, + 0.013759010471403599, + 0.012251787818968296, + 0.01119404286146164, + 0.007742763962596655, + 0.007190405391156673, + 0.02387593686580658, + 0.030506039038300514, + 0.029771024361252785, + 0.04439159110188484, + 0.0228104367852211, + 0.03120485320687294, + 0.017776528373360634, + 0.010211342945694923, + 0.0061041624285280704, + 0.012544078752398491, + 0.010141180828213692, + 0.011218794621527195, + 0.00774629320949316, + 0.00741229485720396, + 0.0046754442155361176, + 0.002939971163868904, + 0.003917127847671509, + 0.00874622818082571, + 0.005612233653664589, + 0.007299846038222313, + 0.006607562303543091, + 0.006523770745843649, + 0.003322963137179613, + 0.005599572788923979, + 0.002862663008272648, + 0.007588847540318966, + 0.00562117388471961, + 0.0072929514572024345, + 0.004848008509725332, + 0.00551340077072382, + 0.002680405043065548, + 0.002769265789538622, + 0.0023976133670657873, + 0.005426550749689341, + 0.004679806064814329, + 0.006319760344922543, + 0.005539681296795607, + 0.004351936746388674, + 0.002663198160007596, + 0.0028099361807107925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.014196357689797878 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428130567073822 + ], + "opposite_logprob_tokens": [ + -2.017812967300415 + ], + "logprob_reward": -0.1428130567073822, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1428130567073822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14371836185455322, + 0.08020473271608353, + 0.16043107211589813, + 0.09141554683446884, + 0.11240204423666, + 0.1666838675737381, + 0.0866205170750618, + 0.09507152438163757, + 0.026429159566760063, + 0.03702316805720329 + ], + "credit_map": [ + 0.012156043201684952, + 0.026733551174402237, + 0.02040827088057995, + 0.021111635491251945, + 0.017786886543035507, + 0.019695108756422997, + 0.012211690656840801, + 0.013615196570754051, + 0.007128924131393433, + 0.015773089602589607, + 0.00970672070980072, + 0.013233311474323273, + 0.009091930463910103, + 0.011626183986663818, + 0.007699948735535145, + 0.005944623611867428, + 0.014823702163994312, + 0.027364715933799744, + 0.015493519604206085, + 0.031103795394301414, + 0.027047298848628998, + 0.019057344645261765, + 0.01747630536556244, + 0.008064385503530502, + 0.007263109087944031, + 0.0112915663048625, + 0.014563544653356075, + 0.01733238622546196, + 0.008656785823404789, + 0.012940685264766216, + 0.010174931958317757, + 0.009192537516355515, + 0.009184475056827068, + 0.024455975741147995, + 0.013471589423716068, + 0.018257269635796547, + 0.017119744792580605, + 0.012276249937713146, + 0.011319557204842567, + 0.006317178253084421, + 0.015629351139068604, + 0.0244005024433136, + 0.02232317067682743, + 0.03722374513745308, + 0.020244784653186798, + 0.019255217164754868, + 0.01714167930185795, + 0.010465428233146667, + 0.007277665194123983, + 0.016231238842010498, + 0.01330562960356474, + 0.015336094424128532, + 0.007976120337843895, + 0.011146134696900845, + 0.0061951205134391785, + 0.009152520447969437, + 0.006453376729041338, + 0.01768593303859234, + 0.010285725817084312, + 0.013465243391692638, + 0.010868643410503864, + 0.01522857416421175, + 0.007533533498644829, + 0.013550498522818089, + 0.0023615665268152952, + 0.004316389560699463, + 0.00458476273342967, + 0.004735629074275494, + 0.0022983180824667215, + 0.0037457547150552273, + 0.0023238782305270433, + 0.002062863437458873, + 0.0031008748337626457, + 0.0050965240225195885, + 0.005072519648820162, + 0.0068823955953121185, + 0.005878675729036331, + 0.005272306501865387, + 0.0030230775009840727, + 0.002696795854717493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.1428130567073822 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a horizontal angle, the image exudes Rembrandt-like light and shadow, reminiscent of a classical oil painting, its brushstrokes intense and dense. The sole illumination comes from a flickering candlelight in the lower center of the frame. This thick, half-burned candle rests in a rustic clay candlestick, its warm orange glow the centerpiece of the dim scene. The light gently illuminates the face of an elderly woman, whose deep wrinkles, like a dry riverbed, crisscross across her forehead, eyes, and cheeks, each bearing a story. Her silver-white hair is loosely tied back with a simple wooden hairpin, with a few strands falling by her cheeks. She wears a dark brown, rough linen top. She leans slightly forward, her cloudy yet luminous eyes fixed intently on the dancing flame, her lips moving at an almost inaudible rhythm, as if engaged in a private conversation with time or a deity. The candlelight casts vivid shadows on her deep wrinkles, trembling slightly with the flickering flame, while the background is completely swallowed by the deep and indissoluble darkness, creating an atmosphere of tranquility, solemnity, contemplation and mystery.", + "seed": 26869, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_13/index_3856_seed_26869_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.172915692999959, + "vlm_logprob": -0.172915692999959, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.172915692999959, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01670151948928833 + ], + "opposite_logprob_tokens": [ + -4.266701698303223 + ], + "logprob_reward": -0.01670151948928833, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01670151948928833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.425517737865448, + 0.0643322765827179, + 0.051687926054000854, + 0.0887104943394661, + 0.037390630692243576, + 0.14309905469417572, + 0.060124244540929794, + 0.04269637167453766, + 0.06668777763843536, + 0.019753502681851387 + ], + "credit_map": [ + 0.034541934728622437, + 0.04049471765756607, + 0.08850521594285965, + 0.0497664250433445, + 0.05567839369177818, + 0.08114095032215118, + 0.05783996358513832, + 0.017550131306052208, + 0.006053904537111521, + 0.005444453563541174, + 0.015269650146365166, + 0.005626644007861614, + 0.005506867542862892, + 0.0173802450299263, + 0.005390403792262077, + 0.0036601058673113585, + 0.0048987725749611855, + 0.005297009367495775, + 0.01056464109569788, + 0.004037010483443737, + 0.004998978227376938, + 0.011982588097453117, + 0.0034525974187999964, + 0.006456328555941582, + 0.004132951609790325, + 0.005697058048099279, + 0.029907923191785812, + 0.004905949346721172, + 0.004082110710442066, + 0.032470330595970154, + 0.0028360318392515182, + 0.004678140394389629, + 0.0028826184570789337, + 0.005278321914374828, + 0.008116863667964935, + 0.005603353958576918, + 0.003937980160117149, + 0.0054195113480091095, + 0.004379033111035824, + 0.0017729507526382804, + 0.011934333480894566, + 0.0179420355707407, + 0.029742851853370667, + 0.017967022955417633, + 0.017013609409332275, + 0.021070539951324463, + 0.018369656056165695, + 0.009059005416929722, + 0.0074867685325443745, + 0.0086336899548769, + 0.010118265636265278, + 0.007201968226581812, + 0.008465343154966831, + 0.007838847115635872, + 0.006318176630884409, + 0.004061184357851744, + 0.005155261605978012, + 0.00620841421186924, + 0.009082216769456863, + 0.0039366791024804115, + 0.005128184799104929, + 0.006603310815989971, + 0.00434125168249011, + 0.0022410533856600523, + 0.0051316795870661736, + 0.00822574831545353, + 0.010244647040963173, + 0.011803784407675266, + 0.006903265602886677, + 0.010868995450437069, + 0.008678383193910122, + 0.00483127124607563, + 0.0023098725359886885, + 0.0021029782947152853, + 0.0030482870060950518, + 0.003387375269085169, + 0.0026422198861837387, + 0.0028863269835710526, + 0.0017551254713907838, + 0.001621317584067583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.01670151948928833 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820521429181099 + ], + "opposite_logprob_tokens": [ + -4.018205165863037 + ], + "logprob_reward": -0.01820521429181099, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820521429181099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13517040014266968, + 0.0780424252152443, + 0.06097402796149254, + 0.2393949031829834, + 0.06215736269950867, + 0.044541265815496445, + 0.09527712315320969, + 0.08703490346670151, + 0.08591847866773605, + 0.11148912459611893 + ], + "credit_map": [ + 0.014849490486085415, + 0.014158464968204498, + 0.022329937666654587, + 0.015818316489458084, + 0.016532663255929947, + 0.01871453784406185, + 0.012731221504509449, + 0.020035769790410995, + 0.008234215900301933, + 0.013105826452374458, + 0.010963601991534233, + 0.007526555098593235, + 0.008476763032376766, + 0.010949039831757545, + 0.0077443309128284454, + 0.011042091995477676, + 0.0052286661230027676, + 0.006435368210077286, + 0.011960810981690884, + 0.005700033623725176, + 0.0077979364432394505, + 0.013479769229888916, + 0.0065433247946202755, + 0.003828121582046151, + 0.01388462632894516, + 0.008072909899055958, + 0.08570334315299988, + 0.013003461062908173, + 0.007050069514662027, + 0.08994291722774506, + 0.007110214792191982, + 0.014627360738813877, + 0.006499222479760647, + 0.006686186883598566, + 0.01105455867946148, + 0.00777185196056962, + 0.007062346208840609, + 0.009473418816924095, + 0.008507235907018185, + 0.005102545488625765, + 0.005309853702783585, + 0.005149646196514368, + 0.0061104921624064445, + 0.00468668295070529, + 0.007343894336372614, + 0.0055263894610106945, + 0.006706364918500185, + 0.003707943484187126, + 0.008860094472765923, + 0.013066303916275501, + 0.015764137730002403, + 0.012867606244981289, + 0.015631072223186493, + 0.014001297764480114, + 0.01111240591853857, + 0.003974201623350382, + 0.012489267624914646, + 0.010316831059753895, + 0.014645403251051903, + 0.009220287203788757, + 0.012300382368266582, + 0.01200874987989664, + 0.009528722614049911, + 0.006525262724608183, + 0.009158071130514145, + 0.00852653943002224, + 0.016314595937728882, + 0.011950021609663963, + 0.008493412286043167, + 0.013708066195249557, + 0.01049620658159256, + 0.007271569687873125, + 0.011598856188356876, + 0.011415100656449795, + 0.016049301251769066, + 0.01675202138721943, + 0.00730234244838357, + 0.017828037962317467, + 0.00899033434689045, + 0.021553127095103264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.01820521429181099 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062070805579423904 + ], + "opposite_logprob_tokens": [ + -2.812070846557617 + ], + "logprob_reward": -0.062070805579423904, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062070805579423904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17831632494926453, + 0.06685397028923035, + 0.04512688145041466, + 0.25806954503059387, + 0.029076697304844856, + 0.09934575110673904, + 0.07203707844018936, + 0.1518300175666809, + 0.04682547226548195, + 0.05251820385456085 + ], + "credit_map": [ + 0.01742323860526085, + 0.020630693063139915, + 0.029831185936927795, + 0.01918426901102066, + 0.022392764687538147, + 0.025767408311367035, + 0.02025814913213253, + 0.022828612476587296, + 0.010825005359947681, + 0.011495588347315788, + 0.00985733512789011, + 0.006576854735612869, + 0.004385525360703468, + 0.00839157309383154, + 0.006610233336687088, + 0.008711851201951504, + 0.003918965347111225, + 0.005342894233763218, + 0.009867948479950428, + 0.0057433671317994595, + 0.004652114585042, + 0.00649206293746829, + 0.004396478645503521, + 0.004713047761470079, + 0.012584920972585678, + 0.01260250061750412, + 0.09152746200561523, + 0.012459413148462772, + 0.009328718297183514, + 0.09776933491230011, + 0.006896102335304022, + 0.014901107177138329, + 0.0023580051492899656, + 0.0029497353825718164, + 0.005324780009686947, + 0.002653412986546755, + 0.0037257259245961905, + 0.004277184139937162, + 0.0030868914909660816, + 0.004700964782387018, + 0.01305970549583435, + 0.012840598821640015, + 0.02069350704550743, + 0.011434227228164673, + 0.008472518995404243, + 0.015981601551175117, + 0.01013497356325388, + 0.006728621665388346, + 0.008445930667221546, + 0.006371459923684597, + 0.015919001772999763, + 0.01003608014434576, + 0.008189987391233444, + 0.011218668892979622, + 0.006734009366482496, + 0.005121944472193718, + 0.015304703265428543, + 0.02390773966908455, + 0.02759346179664135, + 0.016979187726974487, + 0.025715315714478493, + 0.021244673058390617, + 0.011953589506447315, + 0.009131360799074173, + 0.005750781856477261, + 0.006523254327476025, + 0.006536752451211214, + 0.007437269669026136, + 0.00397154688835144, + 0.00732708303257823, + 0.005143257789313793, + 0.004135523922741413, + 0.006224846467375755, + 0.009867442771792412, + 0.006078771781176329, + 0.007705303840339184, + 0.0062245395965874195, + 0.007639733608812094, + 0.005116504617035389, + 0.003661063499748707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.062070805579423904 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097829420119524 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.0097829420119524, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097829420119524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1833433210849762, + 0.06944287568330765, + 0.0484054796397686, + 0.2610526382923126, + 0.046551842242479324, + 0.08676596730947495, + 0.0464903898537159, + 0.1063019186258316, + 0.04460986331105232, + 0.10703576356172562 + ], + "credit_map": [ + 0.014980209991335869, + 0.014700350351631641, + 0.029090890660881996, + 0.021473782137036324, + 0.02486627735197544, + 0.031084783375263214, + 0.028266876935958862, + 0.018880149349570274, + 0.008248688653111458, + 0.013068265281617641, + 0.010962400585412979, + 0.009909553453326225, + 0.005926704965531826, + 0.009543313644826412, + 0.007638146635144949, + 0.0041457959450781345, + 0.006832418497651815, + 0.009217857383191586, + 0.007620120886713266, + 0.00514476140961051, + 0.004198198206722736, + 0.006839756388217211, + 0.00398382730782032, + 0.004568534437566996, + 0.017097748816013336, + 0.012300739996135235, + 0.08934198319911957, + 0.015599217265844345, + 0.007693198975175619, + 0.09592309594154358, + 0.008213737979531288, + 0.01488290075212717, + 0.004602848086506128, + 0.006488924380391836, + 0.00935311894863844, + 0.006025510840117931, + 0.004904851783066988, + 0.006696098949760199, + 0.004773045424371958, + 0.0037074421998113394, + 0.011336960829794407, + 0.011339318938553333, + 0.019200623035430908, + 0.010774013586342335, + 0.008943727239966393, + 0.0142026636749506, + 0.007548843510448933, + 0.003419807879254222, + 0.005216766148805618, + 0.006746688392013311, + 0.010216234251856804, + 0.006132925860583782, + 0.006038196850568056, + 0.006092878989875317, + 0.004065880086272955, + 0.0019808122888207436, + 0.009708145633339882, + 0.011103897355496883, + 0.01579107902944088, + 0.013502351008355618, + 0.01601981557905674, + 0.015277356840670109, + 0.01557593047618866, + 0.009323332458734512, + 0.00528180506080389, + 0.0055740573443472385, + 0.006522693205624819, + 0.006110948510468006, + 0.00394618371501565, + 0.007048757281154394, + 0.005956603214144707, + 0.004168811719864607, + 0.01374478917568922, + 0.023321768268942833, + 0.013685468584299088, + 0.015471206977963448, + 0.010230661369860172, + 0.014514733105897903, + 0.00960458256304264, + 0.006462541408836842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.0097829420119524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated primarily by a flickering candlelight in the lower center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have silver-white hair tied back with a simple wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark brown, rough linen top?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely dark and indistinct?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578179836273193 + ], + "opposite_logprob_tokens": [ + -0.6328179836273193 + ], + "logprob_reward": -0.7578179836273193, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578179836273193, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09187834709882736, + 0.06610748916864395, + 0.10881594568490982, + 0.21179640293121338, + 0.022999752312898636, + 0.014115575700998306, + 0.04988659545779228, + 0.07730015367269516, + 0.2769107222557068, + 0.0801890566945076 + ], + "credit_map": [ + 0.01214446872472763, + 0.012736424803733826, + 0.014481978490948677, + 0.012691721320152283, + 0.015176585875451565, + 0.011235527694225311, + 0.008863134309649467, + 0.004548503551632166, + 0.008799147792160511, + 0.010816682130098343, + 0.010112538933753967, + 0.0072431727312505245, + 0.006186964455991983, + 0.010416505858302116, + 0.006735913921147585, + 0.005796562414616346, + 0.013892433606088161, + 0.012823537923395634, + 0.021520471200346947, + 0.01149984821677208, + 0.009750588797032833, + 0.02417735569179058, + 0.008384499698877335, + 0.006767206825315952, + 0.012205025181174278, + 0.014054130762815475, + 0.06942543387413025, + 0.012166055850684643, + 0.010183539241552353, + 0.07400959730148315, + 0.008788049221038818, + 0.010964564979076385, + 0.0019098573829978704, + 0.003303531091660261, + 0.004167383536696434, + 0.003079259768128395, + 0.002442264696583152, + 0.003153824945911765, + 0.0032094409689307213, + 0.001734189921990037, + 0.0012608339311555028, + 0.0019535678438842297, + 0.0030889399349689484, + 0.0016809324733912945, + 0.0022501489147543907, + 0.0020480623934417963, + 0.0011228009825572371, + 0.0007102885865606368, + 0.004350879229605198, + 0.004625523462891579, + 0.009174160659313202, + 0.006842991337180138, + 0.006704716477543116, + 0.007932409644126892, + 0.006382678635418415, + 0.0038732343818992376, + 0.01019678357988596, + 0.013292236253619194, + 0.012505450285971165, + 0.009558821097016335, + 0.007643772754818201, + 0.012132708914577961, + 0.005887345876544714, + 0.006083029787987471, + 0.033520691096782684, + 0.032117340713739395, + 0.052754223346710205, + 0.04763910546898842, + 0.021513182669878006, + 0.04936542361974716, + 0.026483476161956787, + 0.013517238199710846, + 0.01215587742626667, + 0.013594642281532288, + 0.010944058187305927, + 0.01359642669558525, + 0.009866217151284218, + 0.01070806011557579, + 0.005620091687887907, + 0.0037036852445453405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.7578179836273193 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_14/index_tensor([5217, 5217])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_14/index_tensor([5217, 5217])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c80146380673a383b8ff5d048745f0e65eb78d69 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_14/index_tensor([5217, 5217])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle tracking shot smoothly follows a massive, lifelike watermelon as it scampers across a vast, sunny meadow with astonishing agility. The camera remains low to the ground, emphasizing the fruit's enormous size, the size of a small boulder. Its striped dark green and pale lemon green rind has a surreal, waxy sheen, with every subtle bump and imperfection rendered in razor-sharp detail, gleaming in the direct afternoon sunlight. From its base emerge two thick, exhausted vine legs that churn and pump vigorously, kicking up tiny leaves with each powerful stride. These newly formed branches are a vibrant, living green, bearing small, tender leaves and curling tendrils that flutter in the air. The watermelon's immense weight causes its body to subtly sway in a determined, rhythmic motion. In the background, a clear, bright blue sky stretches to the distant, soft-focus treeline, creating a sense of open, idyllic space. The entire scene is bathed in the warm golden light of the afternoon, casting long, soft shadows behind the running fruits, creating a vivid atmosphere that is surreal, whimsical, yet strikingly vivid.", + "seed": 24746, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_14/index_5217_seed_24746_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.00450761909596622, + "vlm_logprob": -0.00450761909596622, + "vlm_yes_no_margin": 5.775, + "vlm_token_credit": -0.00450761909596622, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003280022880062461 + ], + "opposite_logprob_tokens": [ + -6.628280162811279 + ], + "logprob_reward": -0.003280022880062461, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.003280022880062461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12543949484825134, + 0.043817851692438126, + 0.07043764740228653, + 0.056192461401224136, + 0.27769774198532104, + 0.08492273837327957, + 0.19165967404842377, + 0.056272853165864944, + 0.03836992383003235, + 0.055189620703458786 + ], + "credit_map": [ + 0.01910407468676567, + 0.010803932324051857, + 0.024690307676792145, + 0.00743780005723238, + 0.010379084385931492, + 0.018126187846064568, + 0.01202306617051363, + 0.022875044494867325, + 0.005933769978582859, + 0.005750523880124092, + 0.004780550021678209, + 0.004207344260066748, + 0.004809718579053879, + 0.005084161181002855, + 0.004730416927486658, + 0.008521368727087975, + 0.007813205011188984, + 0.010720694437623024, + 0.007980787195265293, + 0.008526423946022987, + 0.009817440994083881, + 0.009171659126877785, + 0.007880028337240219, + 0.008527407422661781, + 0.006099514197558165, + 0.006157253868877888, + 0.00907090213149786, + 0.00582941435277462, + 0.008294274099171162, + 0.004063654690980911, + 0.008068914525210857, + 0.008608534932136536, + 0.028793806210160255, + 0.021631337702274323, + 0.08673275262117386, + 0.07762827724218369, + 0.01901927962899208, + 0.010011818259954453, + 0.018171027302742004, + 0.015709463506937027, + 0.005372802726924419, + 0.0030857303645461798, + 0.030426811426877975, + 0.007490137126296759, + 0.0038990918546915054, + 0.003355956170707941, + 0.005096291191875935, + 0.026195917278528214, + 0.01441424060612917, + 0.010390421375632286, + 0.023364363238215446, + 0.05377050116658211, + 0.030908770859241486, + 0.010529802180826664, + 0.02442331425845623, + 0.023858262225985527, + 0.010076935403048992, + 0.0072189196944236755, + 0.008471284061670303, + 0.007079842034727335, + 0.004803136456757784, + 0.004483988042920828, + 0.0061888061463832855, + 0.007949940860271454, + 0.0057618883438408375, + 0.004517777822911739, + 0.005034289788454771, + 0.0045536356046795845, + 0.006116548553109169, + 0.0028516962192952633, + 0.005797246005386114, + 0.0037368445191532373, + 0.007426938973367214, + 0.006836472544819117, + 0.009397330693900585, + 0.007703469600528479, + 0.006950243841856718, + 0.006066197529435158, + 0.005611996166408062, + 0.005196969490498304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.003280022880062461 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090893298387527 + ], + "opposite_logprob_tokens": [ + -4.5110907554626465 + ], + "logprob_reward": -0.011090893298387527, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090893298387527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0634990930557251, + 0.04851754754781723, + 0.031411461532115936, + 0.0561218336224556, + 0.209811732172966, + 0.09032447636127472, + 0.338612824678421, + 0.04902707040309906, + 0.07279109954833984, + 0.039882905781269073 + ], + "credit_map": [ + 0.01266210526227951, + 0.0071106282994151115, + 0.011809991672635078, + 0.005010056309401989, + 0.004037384409457445, + 0.007804388180375099, + 0.004455442074686289, + 0.010609098710119724, + 0.0053351703099906445, + 0.008199588395655155, + 0.005550202913582325, + 0.005557722877711058, + 0.003771450836211443, + 0.007932747714221478, + 0.004432754125446081, + 0.00773791316896677, + 0.003486515721306205, + 0.0040216813795268536, + 0.003082876093685627, + 0.004054795019328594, + 0.0042455182410776615, + 0.0052465880289673805, + 0.0033815240021795034, + 0.0038919609505683184, + 0.0063394601456820965, + 0.005106344353407621, + 0.006238649599254131, + 0.006660171318799257, + 0.006768302991986275, + 0.008477363735437393, + 0.007440032437443733, + 0.009091510437428951, + 0.01910582184791565, + 0.018529832363128662, + 0.06279288977384567, + 0.06132293492555618, + 0.013509983196854591, + 0.00796909723430872, + 0.012689766474068165, + 0.013891389593482018, + 0.005066249053925276, + 0.004260745365172625, + 0.028527047485113144, + 0.009957205504179, + 0.0057809678837656975, + 0.003785424865782261, + 0.006456136703491211, + 0.026490693911910057, + 0.02245824970304966, + 0.02452649176120758, + 0.04982677474617958, + 0.10577836632728577, + 0.02692275308072567, + 0.019074734300374985, + 0.04072361811995506, + 0.04930185526609421, + 0.0055210404098033905, + 0.005140654742717743, + 0.00639795558527112, + 0.006223930511623621, + 0.007589282002300024, + 0.004780404735356569, + 0.007474533747881651, + 0.0058992719277739525, + 0.0065252468921244144, + 0.005499774590134621, + 0.013209041208028793, + 0.010080340318381786, + 0.01380365714430809, + 0.00463139871135354, + 0.01206419337540865, + 0.006977440789341927, + 0.004275070037692785, + 0.00430492777377367, + 0.004768981132656336, + 0.005329026374965906, + 0.00402202969416976, + 0.0046595241874456406, + 0.007506370544433594, + 0.005016972776502371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.011090893298387527 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031923307105898857 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031923307105898857, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031923307105898857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05889704078435898, + 0.11016885191202164, + 0.1498410552740097, + 0.04143726825714111, + 0.1946546733379364, + 0.039092015475034714, + 0.18609118461608887, + 0.08929421007633209, + 0.10406772792339325, + 0.026455946266651154 + ], + "credit_map": [ + 0.009888019412755966, + 0.007042400538921356, + 0.010330874472856522, + 0.0059596337378025055, + 0.004330083262175322, + 0.006567683536559343, + 0.005901622120290995, + 0.00887672696262598, + 0.008461062796413898, + 0.022564252838492393, + 0.010733263567090034, + 0.011148693040013313, + 0.015359229408204556, + 0.013776140287518501, + 0.013125970959663391, + 0.015000240877270699, + 0.01510537602007389, + 0.022197632119059563, + 0.012993039563298225, + 0.018087025731801987, + 0.019917894154787064, + 0.02363571897149086, + 0.020846888422966003, + 0.01705748587846756, + 0.004816378466784954, + 0.0033244674559682608, + 0.00568730803206563, + 0.004830888472497463, + 0.004199679009616375, + 0.005355582572519779, + 0.0054376330226659775, + 0.007785330526530743, + 0.014015249907970428, + 0.0128893181681633, + 0.05864971503615379, + 0.06362685561180115, + 0.012557050213217735, + 0.006885870825499296, + 0.012189132161438465, + 0.013841469772160053, + 0.004304914269596338, + 0.003894575871527195, + 0.0043565197847783566, + 0.006799265276640654, + 0.006383213214576244, + 0.003326268633827567, + 0.004426290281116962, + 0.005600967910140753, + 0.01570327766239643, + 0.01667766645550728, + 0.030087491497397423, + 0.043704625219106674, + 0.02253093011677265, + 0.010572989471256733, + 0.023038465529680252, + 0.023775752633810043, + 0.012467901222407818, + 0.009012228809297085, + 0.017370140179991722, + 0.010314421728253365, + 0.014588079415261745, + 0.005882854573428631, + 0.011179005727171898, + 0.008479580283164978, + 0.006671807263046503, + 0.006757762283086777, + 0.01793806627392769, + 0.01042099203914404, + 0.026044996455311775, + 0.0069978609681129456, + 0.018980015069246292, + 0.010256236419081688, + 0.0034981968346983194, + 0.0033888116013258696, + 0.0038376885931938887, + 0.004445836879312992, + 0.0022217659279704094, + 0.002289894036948681, + 0.0025491020642220974, + 0.004224651493132114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.0031923307105898857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024877809919416904 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024877809919416904, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024877809919416904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1018553078174591, + 0.06300147622823715, + 0.16852007806301117, + 0.04940376058220863, + 0.32220491766929626, + 0.061694640666246414, + 0.09066028892993927, + 0.06965301930904388, + 0.02689322642982006, + 0.04611329361796379 + ], + "credit_map": [ + 0.016746342182159424, + 0.014633562415838242, + 0.014634966850280762, + 0.009718167595565319, + 0.008808666840195656, + 0.015162239782512188, + 0.008567371405661106, + 0.013584000989794731, + 0.006209416314959526, + 0.010929657146334648, + 0.0070620509795844555, + 0.005135366227477789, + 0.009480806067585945, + 0.010373087599873543, + 0.006913984660059214, + 0.006897111888974905, + 0.01698923297226429, + 0.028400272130966187, + 0.01855149306356907, + 0.017984094098210335, + 0.023284275084733963, + 0.02671968564391136, + 0.015839708968997, + 0.020751314237713814, + 0.004411808215081692, + 0.00445108488202095, + 0.005663319490849972, + 0.00511760963127017, + 0.006294440012425184, + 0.004842775408178568, + 0.01097537949681282, + 0.0076473490335047245, + 0.046570662409067154, + 0.01650705747306347, + 0.11953455954790115, + 0.06941233575344086, + 0.015591983683407307, + 0.013036610558629036, + 0.023208845406770706, + 0.018342871218919754, + 0.005136193707585335, + 0.0029609519988298416, + 0.01864834688603878, + 0.006899709813296795, + 0.004000524524599314, + 0.0030692387372255325, + 0.005804602522403002, + 0.015175077132880688, + 0.006231442093849182, + 0.007479816675186157, + 0.01063836645334959, + 0.027773894369602203, + 0.009077714756131172, + 0.005037939175963402, + 0.012039032764732838, + 0.01238208170980215, + 0.011644765734672546, + 0.008527603931725025, + 0.010722644627094269, + 0.006816501263529062, + 0.0083880964666605, + 0.005977824795991182, + 0.009516213089227676, + 0.008059374988079071, + 0.005396912340074778, + 0.004069542977958918, + 0.003597050905227661, + 0.003619043854996562, + 0.002740914700552821, + 0.0018755592172965407, + 0.002892363118007779, + 0.0027018391992896795, + 0.005190061870962381, + 0.006093487609177828, + 0.005245022010058165, + 0.006135274190455675, + 0.004839747212827206, + 0.005609739571809769, + 0.0055472166277468204, + 0.007452750578522682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0024877809919416904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002487067598849535 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.002487067598849535, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002487067598849535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1451893001794815, + 0.24404117465019226, + 0.06773769855499268, + 0.046128202229738235, + 0.2046940177679062, + 0.08772673457860947, + 0.06556850671768188, + 0.029458215460181236, + 0.05677546560764313, + 0.05268068239092827 + ], + "credit_map": [ + 0.030199671164155006, + 0.02188575640320778, + 0.02427920140326023, + 0.012804037891328335, + 0.010614311322569847, + 0.018294058740139008, + 0.011397532187402248, + 0.0157147329300642, + 0.03427228331565857, + 0.014448442496359348, + 0.02176916040480137, + 0.013578387908637524, + 0.012000009417533875, + 0.07871278375387192, + 0.014324993826448917, + 0.05493510141968727, + 0.005767976399511099, + 0.010847528465092182, + 0.007951639592647552, + 0.00939925480633974, + 0.009138842113316059, + 0.007366944570094347, + 0.008648167364299297, + 0.008617346175014973, + 0.006066870875656605, + 0.0029144047293812037, + 0.005976845510303974, + 0.005599447526037693, + 0.005574095994234085, + 0.004371219780296087, + 0.009700624272227287, + 0.005924691911786795, + 0.029037337750196457, + 0.0201644916087389, + 0.06135917082428932, + 0.045481447130441666, + 0.01317104883491993, + 0.009535697288811207, + 0.013103462755680084, + 0.012841351330280304, + 0.005985484458506107, + 0.0034595495089888573, + 0.03179953619837761, + 0.008001395501196384, + 0.0028734563384205103, + 0.0032771769911050797, + 0.005373772699385881, + 0.026956364512443542, + 0.005160127300769091, + 0.005149956792593002, + 0.010346464812755585, + 0.017628604546189308, + 0.006920821964740753, + 0.00392558379098773, + 0.008155040442943573, + 0.008281905204057693, + 0.0027881981804966927, + 0.0030297490302473307, + 0.003609756240621209, + 0.003632779000326991, + 0.0050183795392513275, + 0.003363832365721464, + 0.0043802410364151, + 0.003635282628238201, + 0.006201899144798517, + 0.005463429726660252, + 0.009250612929463387, + 0.0044150748290121555, + 0.011636417359113693, + 0.0038562961854040623, + 0.010497779585421085, + 0.005453960038721561, + 0.005400652997195721, + 0.006737617310136557, + 0.006775552406907082, + 0.006846583914011717, + 0.007494388148188591, + 0.004318437539041042, + 0.006861200090497732, + 0.00824624765664339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.002487067598849535 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot smoothly follows a massive, lifelike watermelon as it scampers across a vast, sunny meadow with astonishing agility. The camera remains low to the ground, emphasizing the fruit's enormous size, the size of a small boulder. Its striped dark green and pale lemon green rind has a surreal, waxy sheen, with every subtle bump and imperfection rendered in razor-sharp detail, gleaming in the direct afternoon sunlight. From its base emerge two thick, exhausted vine legs that churn and pump vigorously, kicking up tiny leaves with each powerful stride. These newly formed branches are a vibrant, living green, bearing small, tender leaves and curling tendrils that flutter in the air. The watermelon's immense weight causes its body to subtly sway in a determined, rhythmic motion. In the background, a clear, bright blue sky stretches to the distant, soft-focus treeline, creating a sense of open, idyllic space. The entire scene is bathed in the warm golden light of the afternoon, casting long, soft shadows behind the running fruits, creating a vivid atmosphere that is surreal, whimsical, yet strikingly vivid.", + "seed": 24747, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_14/index_5217_seed_24747_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008070729114115239, + "vlm_logprob": -0.008070729114115239, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.008070729114115239, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059786769561469555 + ], + "opposite_logprob_tokens": [ + -6.630978584289551 + ], + "logprob_reward": -0.0059786769561469555, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0059786769561469555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15204453468322754, + 0.05011482536792755, + 0.058861393481492996, + 0.06725184619426727, + 0.12407172471284866, + 0.06956211477518082, + 0.04775838181376457, + 0.048547662794589996, + 0.2931582033634186, + 0.0886293277144432 + ], + "credit_map": [ + 0.01052098162472248, + 0.03568477928638458, + 0.014287487603724003, + 0.02615831419825554, + 0.0261179581284523, + 0.010808080434799194, + 0.01680932566523552, + 0.011657602153718472, + 0.004908709321171045, + 0.010915914550423622, + 0.006372939795255661, + 0.007732533849775791, + 0.0058667720295488834, + 0.003688465105369687, + 0.0039050430059432983, + 0.006724443752318621, + 0.0049691214226186275, + 0.009807618334889412, + 0.012455850839614868, + 0.005971121601760387, + 0.0036936765536665916, + 0.008547169156372547, + 0.004128869157284498, + 0.00928796548396349, + 0.007147188764065504, + 0.02049865573644638, + 0.007276551332324743, + 0.00828059297055006, + 0.008141870610415936, + 0.0046964120119810104, + 0.006555316038429737, + 0.004655254539102316, + 0.01844952069222927, + 0.014406176283955574, + 0.05548301339149475, + 0.016689132899045944, + 0.002740847645327449, + 0.003192210802808404, + 0.004864542745053768, + 0.008246275596320629, + 0.004639213439077139, + 0.019518796354532242, + 0.02007957547903061, + 0.004593078978359699, + 0.006031754892319441, + 0.0038528922013938427, + 0.005722649861127138, + 0.005124157294631004, + 0.004351919516921043, + 0.006081752013415098, + 0.006055688485503197, + 0.0077414195984601974, + 0.008734092116355896, + 0.004936686251312494, + 0.005487813148647547, + 0.004369010683149099, + 0.004575779661536217, + 0.01276242546737194, + 0.004048242699354887, + 0.006911770906299353, + 0.006528008263558149, + 0.0038234847597777843, + 0.0047602904960513115, + 0.005137663334608078, + 0.025896497070789337, + 0.06467422842979431, + 0.0430622436106205, + 0.03602138161659241, + 0.037055980414152145, + 0.03648941218852997, + 0.023904601112008095, + 0.02605385333299637, + 0.007531163282692432, + 0.01688150130212307, + 0.01473611406981945, + 0.015547447837889194, + 0.01185668259859085, + 0.006522211246192455, + 0.00841397326439619, + 0.007140233181416988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0059786769561469555 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020600058138370514 + ], + "opposite_logprob_tokens": [ + -3.8956000804901123 + ], + "logprob_reward": -0.020600058138370514, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020600058138370514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10375069826841354, + 0.04899132624268532, + 0.055566754192113876, + 0.05695399269461632, + 0.39244544506073, + 0.07807877659797668, + 0.08065047115087509, + 0.07059618085622787, + 0.051883406937122345, + 0.06108295917510986 + ], + "credit_map": [ + 0.008625245653092861, + 0.02510838955640793, + 0.011057259514927864, + 0.014369433745741844, + 0.017198868095874786, + 0.012021860107779503, + 0.007639194373041391, + 0.007730452809482813, + 0.0037910817191004753, + 0.009422818198800087, + 0.006448843516409397, + 0.007988513447344303, + 0.00480604637414217, + 0.005136966705322266, + 0.005601478274911642, + 0.00579557940363884, + 0.004955193493515253, + 0.010260495357215405, + 0.0062419031746685505, + 0.007724291644990444, + 0.0049160802736878395, + 0.010241523385047913, + 0.004034331534057856, + 0.007192935328930616, + 0.00600421754643321, + 0.015045533888041973, + 0.006146362982690334, + 0.006520123686641455, + 0.006320964079350233, + 0.00664104288443923, + 0.005440913140773773, + 0.004834835417568684, + 0.05836556851863861, + 0.047670330852270126, + 0.17805105447769165, + 0.0518374927341938, + 0.007649852428585291, + 0.009052108973264694, + 0.014553045853972435, + 0.025265993550419807, + 0.0047041731886565685, + 0.01532839983701706, + 0.00628076633438468, + 0.005906161852180958, + 0.007073542103171349, + 0.004652724601328373, + 0.008758598007261753, + 0.02537441998720169, + 0.006324206478893757, + 0.012398697435855865, + 0.008533132262527943, + 0.012524940073490143, + 0.015447095967829227, + 0.009245792403817177, + 0.008334815502166748, + 0.007841789163649082, + 0.00567693542689085, + 0.013531872071325779, + 0.006865110248327255, + 0.008612727746367455, + 0.008041319437325, + 0.010379423387348652, + 0.008094610646367073, + 0.009394180029630661, + 0.004761419724673033, + 0.013634461909532547, + 0.008656026795506477, + 0.005327995400875807, + 0.005470368079841137, + 0.005988204386085272, + 0.0035034234169870615, + 0.004541504196822643, + 0.00583632243797183, + 0.013146715238690376, + 0.008540269918739796, + 0.009210068732500076, + 0.008076503872871399, + 0.004647659137845039, + 0.006505042780190706, + 0.005120377521961927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.020600058138370514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031923307105898857 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031923307105898857, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031923307105898857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09025343507528305, + 0.06417947262525558, + 0.035445597022771835, + 0.15260189771652222, + 0.15273921191692352, + 0.13343316316604614, + 0.21704640984535217, + 0.07068073004484177, + 0.05650096759200096, + 0.027119098231196404 + ], + "credit_map": [ + 0.007498928811401129, + 0.023558737710118294, + 0.008425965905189514, + 0.010778929106891155, + 0.014941101893782616, + 0.009957420639693737, + 0.008033146150410175, + 0.007059203926473856, + 0.006031428463757038, + 0.013418138958513737, + 0.007878064177930355, + 0.007607609033584595, + 0.008275330998003483, + 0.006223542615771294, + 0.0060005211271345615, + 0.008744843304157257, + 0.003231604816392064, + 0.0062086558900773525, + 0.00784371793270111, + 0.00442055519670248, + 0.0028048288077116013, + 0.004416000097990036, + 0.0023361684288829565, + 0.004184063524007797, + 0.016787318512797356, + 0.04494181275367737, + 0.016083570197224617, + 0.017985893413424492, + 0.018841901794075966, + 0.012424634769558907, + 0.014264511875808239, + 0.0112722497433424, + 0.021124979481101036, + 0.0188133604824543, + 0.06566254794597626, + 0.01882992498576641, + 0.006064975634217262, + 0.004288859199732542, + 0.006880428176373243, + 0.011074138805270195, + 0.007676393259316683, + 0.03406915441155434, + 0.04756871983408928, + 0.011318831704556942, + 0.010017301887273788, + 0.004522506147623062, + 0.010704409331083298, + 0.007555842399597168, + 0.013799643144011497, + 0.02915441431105137, + 0.022016333416104317, + 0.0418429970741272, + 0.044675618410110474, + 0.019152652472257614, + 0.024963829666376114, + 0.021440912038087845, + 0.005205637775361538, + 0.014083094894886017, + 0.006907672621309757, + 0.0090815220028162, + 0.010107379406690598, + 0.00832466408610344, + 0.008313344791531563, + 0.008657412603497505, + 0.004891586489975452, + 0.011439614929258823, + 0.008917756378650665, + 0.00591946579515934, + 0.006122516933828592, + 0.0077733532525599, + 0.00510686868801713, + 0.0063298046588897705, + 0.002677199197933078, + 0.004596719052642584, + 0.0035567155573517084, + 0.004064285662025213, + 0.0038644096348434687, + 0.0025439728051424026, + 0.002990828361362219, + 0.00282497052103281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.0031923307105898857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046369642950594425 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.0046369642950594425, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046369642950594425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04901253432035446, + 0.05902937427163124, + 0.03789297863841057, + 0.12111250311136246, + 0.03663196414709091, + 0.09032032638788223, + 0.44087594747543335, + 0.0672144889831543, + 0.07998877018690109, + 0.01792108453810215 + ], + "credit_map": [ + 0.004947062581777573, + 0.008316492661833763, + 0.003998470958322287, + 0.010834649205207825, + 0.004984427243471146, + 0.005906180012971163, + 0.004068178590387106, + 0.0059570749290287495, + 0.005883288104087114, + 0.009509168565273285, + 0.008257421664893627, + 0.006421216297894716, + 0.007759805303066969, + 0.007821918465197086, + 0.005588226020336151, + 0.007788330316543579, + 0.005027591250836849, + 0.008715707808732986, + 0.00625374261289835, + 0.00529068149626255, + 0.0025618220679461956, + 0.0029014903120696545, + 0.0029768822714686394, + 0.004165066406130791, + 0.013302045874297619, + 0.035603057593107224, + 0.011477581225335598, + 0.01467989943921566, + 0.015525481663644314, + 0.008884373120963573, + 0.01225495059043169, + 0.009385133162140846, + 0.0031553381122648716, + 0.007221243344247341, + 0.010007212869822979, + 0.003969518933445215, + 0.003474852303043008, + 0.0019342972664162517, + 0.002972072921693325, + 0.0038974308408796787, + 0.004296866711229086, + 0.004993933252990246, + 0.027014927938580513, + 0.0077538397163152695, + 0.006702848710119724, + 0.004479296505451202, + 0.005657434929162264, + 0.029421189799904823, + 0.02598034031689167, + 0.0595732219517231, + 0.04562081769108772, + 0.09765459597110748, + 0.07854888588190079, + 0.035928256809711456, + 0.05469609051942825, + 0.04287378862500191, + 0.005104037467390299, + 0.012763903476297855, + 0.005689393263310194, + 0.00946429930627346, + 0.009459379129111767, + 0.008073282428085804, + 0.008251471444964409, + 0.008408731780946255, + 0.007057126611471176, + 0.017704041674733162, + 0.011290266178548336, + 0.010290367528796196, + 0.009497398510575294, + 0.011867444962263107, + 0.005535351112484932, + 0.006746784318238497, + 0.0016620628302916884, + 0.00198917742818594, + 0.003721748711541295, + 0.0018841512501239777, + 0.0021720502991229296, + 0.0017036391654983163, + 0.0018223662627860904, + 0.00296589033678174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.0046369642950594425 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945615470409393 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.005945615470409393, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945615470409393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1908014416694641, + 0.027834828943014145, + 0.029453594237565994, + 0.022192414849996567, + 0.5055322647094727, + 0.07183808088302612, + 0.04923338443040848, + 0.031056514009833336, + 0.034278977662324905, + 0.03777854144573212 + ], + "credit_map": [ + 0.01495467871427536, + 0.04077974706888199, + 0.016336949542164803, + 0.01954258605837822, + 0.034538596868515015, + 0.016769589856266975, + 0.02614518254995346, + 0.021734092384576797, + 0.002360675483942032, + 0.0035898112691938877, + 0.004384415689855814, + 0.0027146663051098585, + 0.002795392880216241, + 0.0042620631866157055, + 0.0039261747151613235, + 0.003801628015935421, + 0.0023543955758213997, + 0.005064794793725014, + 0.004804679658263922, + 0.002505264710634947, + 0.002032194985076785, + 0.005384763702750206, + 0.0018394171493127942, + 0.005468083545565605, + 0.0024987380020320415, + 0.005632230080664158, + 0.002204383024945855, + 0.003110351739451289, + 0.0021033016964793205, + 0.0015067688655108213, + 0.0024652197025716305, + 0.002671420807018876, + 0.0807182565331459, + 0.04846031218767166, + 0.2285732477903366, + 0.07430341094732285, + 0.009504793211817741, + 0.01247362419962883, + 0.01916849985718727, + 0.032330065965652466, + 0.004708749242126942, + 0.018166538327932358, + 0.025279317051172256, + 0.005758130922913551, + 0.004360349383205175, + 0.0041458336636424065, + 0.005070144310593605, + 0.0043490189127624035, + 0.003349082078784704, + 0.007485557347536087, + 0.004875061102211475, + 0.009863430634140968, + 0.006404878571629524, + 0.007575125899165869, + 0.005394936539232731, + 0.004285310860723257, + 0.0030078808777034283, + 0.0056998226791620255, + 0.003310419386252761, + 0.004109921865165234, + 0.004081432241946459, + 0.00322858733125031, + 0.003951115068048239, + 0.003667331300675869, + 0.0029275724664330482, + 0.005283812060952187, + 0.005508589092642069, + 0.004932779353111982, + 0.0036163527984172106, + 0.004068128764629364, + 0.004027282819151878, + 0.003914455883204937, + 0.002932254457846284, + 0.005240737460553646, + 0.008987409994006157, + 0.0044554914347827435, + 0.004323570057749748, + 0.0033022023271769285, + 0.003933263942599297, + 0.004603611305356026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.005945615470409393 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot smoothly follows a massive, lifelike watermelon as it scampers across a vast, sunny meadow with astonishing agility. The camera remains low to the ground, emphasizing the fruit's enormous size, the size of a small boulder. Its striped dark green and pale lemon green rind has a surreal, waxy sheen, with every subtle bump and imperfection rendered in razor-sharp detail, gleaming in the direct afternoon sunlight. From its base emerge two thick, exhausted vine legs that churn and pump vigorously, kicking up tiny leaves with each powerful stride. These newly formed branches are a vibrant, living green, bearing small, tender leaves and curling tendrils that flutter in the air. The watermelon's immense weight causes its body to subtly sway in a determined, rhythmic motion. In the background, a clear, bright blue sky stretches to the distant, soft-focus treeline, creating a sense of open, idyllic space. The entire scene is bathed in the warm golden light of the afternoon, casting long, soft shadows behind the running fruits, creating a vivid atmosphere that is surreal, whimsical, yet strikingly vivid.", + "seed": 24748, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_14/index_5217_seed_24748_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.006189843500033021, + "vlm_logprob": -0.006189843500033021, + "vlm_yes_no_margin": 5.525, + "vlm_token_credit": -0.006189843500033021, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0033033110667020082 + ], + "opposite_logprob_tokens": [ + -7.128303527832031 + ], + "logprob_reward": -0.0033033110667020082, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0033033110667020082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09046956896781921, + 0.0348702073097229, + 0.2524528205394745, + 0.06679095327854156, + 0.11828520148992538, + 0.06289070844650269, + 0.06881007552146912, + 0.13158577680587769, + 0.08109627664089203, + 0.09274837374687195 + ], + "credit_map": [ + 0.0073549808003008366, + 0.0074997772462666035, + 0.011123137548565865, + 0.0072759101167321205, + 0.023734698072075844, + 0.009688728488981724, + 0.015119926072657108, + 0.008672408759593964, + 0.0026191507931798697, + 0.003970740363001823, + 0.0036288013216108084, + 0.0049478281289339066, + 0.005455743055790663, + 0.004618137143552303, + 0.005790614057332277, + 0.003839195007458329, + 0.044489409774541855, + 0.014346973970532417, + 0.09067869186401367, + 0.023680230602622032, + 0.021042320877313614, + 0.01304029580205679, + 0.02743973769247532, + 0.01773517020046711, + 0.0072751338593661785, + 0.010694445110857487, + 0.007722620386630297, + 0.007197882980108261, + 0.009982186369597912, + 0.009807322174310684, + 0.00783410482108593, + 0.006277264561504126, + 0.009450945071876049, + 0.02201845310628414, + 0.010417957790195942, + 0.017913591116666794, + 0.02095518261194229, + 0.01581878773868084, + 0.014405900612473488, + 0.007304386701434851, + 0.00799659825861454, + 0.0111292889341712, + 0.004696153569966555, + 0.007237222511321306, + 0.008429890498518944, + 0.009097130969166756, + 0.0068070837296545506, + 0.007497342769056559, + 0.00791922863572836, + 0.008122803643345833, + 0.012799154967069626, + 0.008781347423791885, + 0.00790532398968935, + 0.007525184191763401, + 0.011345998384058475, + 0.004411037545651197, + 0.00890854187309742, + 0.01830725371837616, + 0.015641575679183006, + 0.012551107443869114, + 0.018925156444311142, + 0.007438417058438063, + 0.03436853364109993, + 0.015445196069777012, + 0.008930865675210953, + 0.01687457226216793, + 0.007063566241413355, + 0.009634430520236492, + 0.020626863464713097, + 0.006388239096850157, + 0.006091882474720478, + 0.00548585643991828, + 0.006505967583507299, + 0.01004276517778635, + 0.02970684878528118, + 0.008576735854148865, + 0.00967591442167759, + 0.004659618251025677, + 0.006416437681764364, + 0.017164085060358047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.0033033110667020082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208345673978329 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.014208345673978329, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208345673978329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09210759401321411, + 0.05116535350680351, + 0.15734095871448517, + 0.057070691138505936, + 0.06813709437847137, + 0.04247714579105377, + 0.07322337478399277, + 0.3088504374027252, + 0.041502125561237335, + 0.10812520235776901 + ], + "credit_map": [ + 0.004903457593172789, + 0.00495799258351326, + 0.00983812939375639, + 0.03005531057715416, + 0.024807220324873924, + 0.004920597653836012, + 0.009997689165174961, + 0.0026271999813616276, + 0.005763514433056116, + 0.005473266821354628, + 0.006113320589065552, + 0.006635746452957392, + 0.00932022463530302, + 0.005222862586379051, + 0.007963827811181545, + 0.004672588314861059, + 0.027782157063484192, + 0.008952443487942219, + 0.05347897484898567, + 0.01434126403182745, + 0.00883630383759737, + 0.006636202801018953, + 0.027903113514184952, + 0.009410487487912178, + 0.006935845594853163, + 0.00992492027580738, + 0.005626107100397348, + 0.007872209884226322, + 0.00700894370675087, + 0.0077076866291463375, + 0.006718028802424669, + 0.005276945885270834, + 0.00556546775624156, + 0.012644918635487556, + 0.0072074891068041325, + 0.010922830551862717, + 0.0128025496378541, + 0.006252977531403303, + 0.00857971515506506, + 0.0041611469350755215, + 0.004240482579916716, + 0.0066576371900737286, + 0.005653688218444586, + 0.005696969572454691, + 0.005249370820820332, + 0.0049756672233343124, + 0.0042960611172020435, + 0.005707271862775087, + 0.007469745818525553, + 0.013478436507284641, + 0.008081389591097832, + 0.012949144467711449, + 0.009307939559221268, + 0.005865500774234533, + 0.012608223594725132, + 0.0034629967994987965, + 0.021744661033153534, + 0.03677842393517494, + 0.03384026512503624, + 0.03165162354707718, + 0.04503510147333145, + 0.017280524596571922, + 0.08476696163415909, + 0.03775288909673691, + 0.004877462983131409, + 0.006241183262318373, + 0.006077643018215895, + 0.00534594152122736, + 0.006356355734169483, + 0.003629432525485754, + 0.004951988812536001, + 0.004022119101136923, + 0.009386402554810047, + 0.018738241866230965, + 0.025943893939256668, + 0.01755838468670845, + 0.012698829174041748, + 0.007707060314714909, + 0.009676875546574593, + 0.006415509153157473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.014208345673978329 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094313830137253 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094313830137253, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094313830137253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11376817524433136, + 0.050307147204875946, + 0.33654579520225525, + 0.0590716190636158, + 0.1268976479768753, + 0.09317667782306671, + 0.035292260348796844, + 0.052874621003866196, + 0.05914202332496643, + 0.07292401045560837 + ], + "credit_map": [ + 0.005314581096172333, + 0.006173240020871162, + 0.009723497554659843, + 0.03979218751192093, + 0.029609274119138718, + 0.005876431707292795, + 0.013456961140036583, + 0.003822007682174444, + 0.005029117222875357, + 0.006801371928304434, + 0.004493799991905689, + 0.004329575691372156, + 0.01049299817532301, + 0.005498667247593403, + 0.007003872189670801, + 0.006657746154814959, + 0.05915067717432976, + 0.019104840233922005, + 0.12034832686185837, + 0.03269215300679207, + 0.021546652540564537, + 0.015955252572894096, + 0.044928379356861115, + 0.022819548845291138, + 0.005299861542880535, + 0.012293443083763123, + 0.00438713701441884, + 0.00884727668017149, + 0.010395566001534462, + 0.007189394906163216, + 0.0063170501962304115, + 0.0043418812565505505, + 0.010788023471832275, + 0.01724962517619133, + 0.0187307670712471, + 0.01570376195013523, + 0.017443202435970306, + 0.02570624276995659, + 0.011449056677520275, + 0.009826958179473877, + 0.007959035225212574, + 0.014280837029218674, + 0.009339326992630959, + 0.011038932017982006, + 0.017160560935735703, + 0.008604076690971851, + 0.010488740168511868, + 0.01430516503751278, + 0.003607823746278882, + 0.006727235391736031, + 0.0043651144951581955, + 0.004576056730002165, + 0.0045292009599506855, + 0.0037742687854915857, + 0.004832506645470858, + 0.0028800524305552244, + 0.006346964742988348, + 0.008022340945899487, + 0.005235607270151377, + 0.00505563011392951, + 0.006790767889469862, + 0.006382280960679054, + 0.009476649574935436, + 0.005564382299780846, + 0.005120833404362202, + 0.00971085112541914, + 0.0073120505549013615, + 0.009083683602511883, + 0.010507497936487198, + 0.0031152518931776285, + 0.004272536374628544, + 0.010019314475357533, + 0.006257866509258747, + 0.015193033032119274, + 0.01002029050141573, + 0.010906104929745197, + 0.01086738146841526, + 0.005747064482420683, + 0.007895887829363346, + 0.006036378908902407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.004094313830137253 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094670061022043 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094670061022043, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094670061022043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11047018319368362, + 0.14741171896457672, + 0.14066843688488007, + 0.08895928412675858, + 0.0793113186955452, + 0.1453390121459961, + 0.058444004505872726, + 0.08819407224655151, + 0.02778615802526474, + 0.11341577768325806 + ], + "credit_map": [ + 0.008810327388346195, + 0.013052302412688732, + 0.009158944711089134, + 0.015789102762937546, + 0.01949021965265274, + 0.016498995944857597, + 0.01814594864845276, + 0.009524344466626644, + 0.01916031911969185, + 0.018771182745695114, + 0.01723640412092209, + 0.01522359810769558, + 0.031045252457261086, + 0.015097959898412228, + 0.0191575326025486, + 0.011719468981027603, + 0.01677628606557846, + 0.009213738143444061, + 0.022267544642090797, + 0.014789649285376072, + 0.023135045543313026, + 0.010043435730040073, + 0.035406991839408875, + 0.009035753086209297, + 0.008019660599529743, + 0.01630260795354843, + 0.00923838373273611, + 0.008654153905808926, + 0.01494180504232645, + 0.014044181443750858, + 0.009679519571363926, + 0.00807897374033928, + 0.0068122996017336845, + 0.010416649281978607, + 0.008735550567507744, + 0.01001051813364029, + 0.0134223448112607, + 0.01521617453545332, + 0.009107794612646103, + 0.005589993670582771, + 0.013882541097700596, + 0.027196869254112244, + 0.011318237520754337, + 0.018409008160233498, + 0.020673589780926704, + 0.01817314140498638, + 0.019336648285388947, + 0.01634897291660309, + 0.007413193583488464, + 0.012181269004940987, + 0.00620419904589653, + 0.009040353819727898, + 0.008015352301299572, + 0.005232103634625673, + 0.007233979646116495, + 0.0031235553324222565, + 0.010833743959665298, + 0.01533716730773449, + 0.010376176796853542, + 0.009984035976231098, + 0.011603832244873047, + 0.007557470351457596, + 0.013906416483223438, + 0.00859522633254528, + 0.0024202330969274044, + 0.004852017853409052, + 0.004171759821474552, + 0.004641834646463394, + 0.004194631706923246, + 0.0025954502634704113, + 0.002537266816943884, + 0.002372962888330221, + 0.006300496868789196, + 0.008204079233109951, + 0.04594918712973595, + 0.01680607534945011, + 0.01163808349519968, + 0.0045936135575175285, + 0.005601502954959869, + 0.014322743751108646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.004094670061022043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248576868325472 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248576868325472, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248576868325472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20401069521903992, + 0.05313251167535782, + 0.05189022794365883, + 0.07917246222496033, + 0.16935613751411438, + 0.07494313269853592, + 0.1028747484087944, + 0.07947679609060287, + 0.0825435146689415, + 0.10259966552257538 + ], + "credit_map": [ + 0.011913764290511608, + 0.011693171225488186, + 0.014194559305906296, + 0.05110815167427063, + 0.06577517837285995, + 0.01585463434457779, + 0.022520821541547775, + 0.010950442403554916, + 0.0060102189891040325, + 0.00766287837177515, + 0.00624928018078208, + 0.005382834002375603, + 0.009955848567187786, + 0.00601537711918354, + 0.007145320065319538, + 0.004710756707936525, + 0.008704452775418758, + 0.0035788994282484055, + 0.01602504961192608, + 0.0048334733583033085, + 0.004340231418609619, + 0.0054388330318033695, + 0.005116961896419525, + 0.0038523331750184298, + 0.009648770093917847, + 0.015983833000063896, + 0.008672329597175121, + 0.010335179045796394, + 0.010272661224007607, + 0.009287807159125805, + 0.008008128963410854, + 0.006963768973946571, + 0.013699292205274105, + 0.03667876869440079, + 0.018012551590800285, + 0.02657245099544525, + 0.025242755189538002, + 0.021010983735322952, + 0.018890883773565292, + 0.009248467162251472, + 0.00943260919302702, + 0.010760383680462837, + 0.008185220882296562, + 0.0098150335252285, + 0.00992544274777174, + 0.008376096375286579, + 0.010728960856795311, + 0.007719389628618956, + 0.0104803666472435, + 0.013247010298073292, + 0.014593098312616348, + 0.014129587449133396, + 0.012427866458892822, + 0.014561137184500694, + 0.01646919548511505, + 0.006966498680412769, + 0.008185872808098793, + 0.01611345447599888, + 0.00978920329362154, + 0.010396817699074745, + 0.010938557796180248, + 0.0051187570206820965, + 0.0111386151984334, + 0.007795520592480898, + 0.005368129815906286, + 0.008715863339602947, + 0.006658877246081829, + 0.014738766476511955, + 0.01548136118799448, + 0.006947976071387529, + 0.006291673518717289, + 0.018340878188610077, + 0.00939990021288395, + 0.014866388402879238, + 0.01109224185347557, + 0.012533202767372131, + 0.015113649889826775, + 0.017188215628266335, + 0.013321110047399998, + 0.009084959514439106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.005248576868325472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot smoothly follows a massive, lifelike watermelon as it scampers across a vast, sunny meadow with astonishing agility. The camera remains low to the ground, emphasizing the fruit's enormous size, the size of a small boulder. Its striped dark green and pale lemon green rind has a surreal, waxy sheen, with every subtle bump and imperfection rendered in razor-sharp detail, gleaming in the direct afternoon sunlight. From its base emerge two thick, exhausted vine legs that churn and pump vigorously, kicking up tiny leaves with each powerful stride. These newly formed branches are a vibrant, living green, bearing small, tender leaves and curling tendrils that flutter in the air. The watermelon's immense weight causes its body to subtly sway in a determined, rhythmic motion. In the background, a clear, bright blue sky stretches to the distant, soft-focus treeline, creating a sense of open, idyllic space. The entire scene is bathed in the warm golden light of the afternoon, casting long, soft shadows behind the running fruits, creating a vivid atmosphere that is surreal, whimsical, yet strikingly vivid.", + "seed": 24749, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_14/index_5217_seed_24749_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.005239991983398795, + "vlm_logprob": -0.005239991983398795, + "vlm_yes_no_margin": 5.475, + "vlm_token_credit": -0.005239991983398795, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036568455398082733 + ], + "opposite_logprob_tokens": [ + -6.25365686416626 + ], + "logprob_reward": -0.0036568455398082733, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0036568455398082733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09564398974180222, + 0.0849851593375206, + 0.23615625500679016, + 0.024275334551930428, + 0.036905787885189056, + 0.04600068926811218, + 0.11623023450374603, + 0.09380103647708893, + 0.19887620210647583, + 0.0671253427863121 + ], + "credit_map": [ + 0.010939964093267918, + 0.013355746865272522, + 0.009146670810878277, + 0.011277232319116592, + 0.02452094666659832, + 0.006279937457293272, + 0.009716116823256016, + 0.010407381691038609, + 0.010113206692039967, + 0.010578124783933163, + 0.008000839501619339, + 0.006487766746431589, + 0.017819153144955635, + 0.010292605496942997, + 0.007878067903220654, + 0.013815387152135372, + 0.01962820254266262, + 0.03646985441446304, + 0.021606704220175743, + 0.04665108397603035, + 0.03051653318107128, + 0.011443630792200565, + 0.03558429330587387, + 0.034255947917699814, + 0.002399359131231904, + 0.002247188938781619, + 0.0042277742177248, + 0.0028894953429698944, + 0.003908168524503708, + 0.0017179002752527595, + 0.0027997924480587244, + 0.004085655324161053, + 0.00417861994355917, + 0.005826930981129408, + 0.004511171020567417, + 0.004137461073696613, + 0.005533761344850063, + 0.004595007747411728, + 0.00415318226441741, + 0.003969650715589523, + 0.004223702009767294, + 0.005223555024713278, + 0.007754760328680277, + 0.004856997635215521, + 0.005435788072645664, + 0.0036246543750166893, + 0.004857576917856932, + 0.010023657232522964, + 0.012458836659789085, + 0.04100346565246582, + 0.006828694138675928, + 0.010590711608529091, + 0.01537954993546009, + 0.014519072137773037, + 0.007607306353747845, + 0.007842594757676125, + 0.006955629214644432, + 0.011224254965782166, + 0.010363786481320858, + 0.013470310717821121, + 0.016158584505319595, + 0.014412499964237213, + 0.011991346254944801, + 0.009224627166986465, + 0.021069718524813652, + 0.03506309911608696, + 0.01073924545198679, + 0.019892936572432518, + 0.04893423244357109, + 0.026209183037281036, + 0.021349016577005386, + 0.015618780627846718, + 0.007718647830188274, + 0.00412813201546669, + 0.0039259870536625385, + 0.004258792381733656, + 0.011668382212519646, + 0.019408518448472023, + 0.005415044259279966, + 0.010601834394037724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.0036568455398082733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791914373636246 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009791914373636246, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791914373636246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15322814881801605, + 0.10735707730054855, + 0.11349048465490341, + 0.02259920723736286, + 0.20362278819084167, + 0.06295771151781082, + 0.07223864644765854, + 0.06196516379714012, + 0.14667493104934692, + 0.05586588382720947 + ], + "credit_map": [ + 0.017598798498511314, + 0.018209131434559822, + 0.015109476633369923, + 0.01906600035727024, + 0.0417976938188076, + 0.011088245548307896, + 0.015143725089728832, + 0.015215055085718632, + 0.009292635135352612, + 0.01606139913201332, + 0.012257738970220089, + 0.010607419535517693, + 0.020387956872582436, + 0.007106045726686716, + 0.014149120077490807, + 0.017494747415184975, + 0.008337191306054592, + 0.01454764511436224, + 0.012239445000886917, + 0.016489652916789055, + 0.014530087821185589, + 0.006220453884452581, + 0.02353781647980213, + 0.017588181421160698, + 0.00202204124070704, + 0.0028646148275583982, + 0.002806013450026512, + 0.004677880555391312, + 0.002817341824993491, + 0.0015344535931944847, + 0.002802741015329957, + 0.003074121428653598, + 0.01788625866174698, + 0.02262517809867859, + 0.017819136381149292, + 0.01809196174144745, + 0.06121712177991867, + 0.01222197711467743, + 0.024695435538887978, + 0.029065707698464394, + 0.004474342335015535, + 0.008978603407740593, + 0.007289322558790445, + 0.011890728957951069, + 0.006866051349788904, + 0.004322133492678404, + 0.009012660942971706, + 0.010123863816261292, + 0.007293326314538717, + 0.023829247802495956, + 0.005458594299852848, + 0.00674777664244175, + 0.009750353172421455, + 0.006562130060046911, + 0.007238729856908321, + 0.005358481779694557, + 0.004380331374704838, + 0.006921209394931793, + 0.011347171850502491, + 0.006973989773541689, + 0.00770460395142436, + 0.005873444490134716, + 0.005583843681961298, + 0.013180562295019627, + 0.012397604063153267, + 0.027557557448744774, + 0.011175457388162613, + 0.02162056602537632, + 0.01969798281788826, + 0.020427197217941284, + 0.01896875537931919, + 0.014829817228019238, + 0.005389038939028978, + 0.007847156375646591, + 0.0037044163327664137, + 0.005129740573465824, + 0.010995364747941494, + 0.006925632711499929, + 0.006922313477844, + 0.008952217176556587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.009791914373636246 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036145609337836504 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036145609337836504, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036145609337836504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18228061497211456, + 0.05328411981463432, + 0.15891343355178833, + 0.017138181254267693, + 0.08528348058462143, + 0.05527357757091522, + 0.08991691470146179, + 0.09926387667655945, + 0.2221258580684662, + 0.03652000427246094 + ], + "credit_map": [ + 0.018261151388287544, + 0.02156769298017025, + 0.017199138179421425, + 0.021451463922858238, + 0.044651053845882416, + 0.014000725001096725, + 0.02220146730542183, + 0.022947916761040688, + 0.00471313064917922, + 0.0044865962117910385, + 0.007648227736353874, + 0.005125260446220636, + 0.00874130055308342, + 0.0030257562175393105, + 0.011846202425658703, + 0.007697643246501684, + 0.01248141285032034, + 0.020396944135427475, + 0.017910892143845558, + 0.02584090083837509, + 0.019792715087532997, + 0.008564440533518791, + 0.031461551785469055, + 0.022464560344815254, + 0.0016115210019052029, + 0.002531385514885187, + 0.0020064020063728094, + 0.0025005079805850983, + 0.003475657431408763, + 0.0012211839202791452, + 0.0017299677710980177, + 0.0020615546964108944, + 0.009683595038950443, + 0.012083613313734531, + 0.009895019233226776, + 0.011945860460400581, + 0.010071485303342342, + 0.005952571053057909, + 0.016632700338959694, + 0.009018640033900738, + 0.004158457275480032, + 0.00886108260601759, + 0.005899837240576744, + 0.010809777304530144, + 0.005945765879005194, + 0.002844284987077117, + 0.00812725443392992, + 0.00862711388617754, + 0.008740781806409359, + 0.02818814478814602, + 0.008988333866000175, + 0.008036346174776554, + 0.010819783434271812, + 0.009300475008785725, + 0.00537113519385457, + 0.01047191396355629, + 0.0069352686405181885, + 0.011944043450057507, + 0.01629495806992054, + 0.026053035631775856, + 0.009409076534211636, + 0.00729446392506361, + 0.01021693553775549, + 0.011116085574030876, + 0.017685187980532646, + 0.04768841341137886, + 0.022011326625943184, + 0.032326389104127884, + 0.026647236198186874, + 0.025600170716643333, + 0.029003486037254333, + 0.021163636818528175, + 0.003573888447135687, + 0.005660464055836201, + 0.002877349965274334, + 0.0030658282339572906, + 0.0066909873858094215, + 0.005212011281400919, + 0.004273729398846626, + 0.005165745038539171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0036145609337836504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945615470409393 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.005945615470409393, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945615470409393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1448284536600113, + 0.05720773711800575, + 0.1174406036734581, + 0.02682085707783699, + 0.04833715409040451, + 0.08874557912349701, + 0.13135387003421783, + 0.10263193398714066, + 0.18557590246200562, + 0.09705787152051926 + ], + "credit_map": [ + 0.015200641937553883, + 0.017100414261221886, + 0.016230622306466103, + 0.019466077908873558, + 0.032293446362018585, + 0.01460878923535347, + 0.016191992908716202, + 0.013736466877162457, + 0.005588498432189226, + 0.00670802965760231, + 0.00676023168489337, + 0.0055041806772351265, + 0.013773963786661625, + 0.006943808402866125, + 0.0055150664411485195, + 0.006413962226361036, + 0.011722736991941929, + 0.01271715760231018, + 0.016048626974225044, + 0.022666847333312035, + 0.010482609272003174, + 0.0047362991608679295, + 0.023384446278214455, + 0.015681887045502663, + 0.0017089175526052713, + 0.002391361864283681, + 0.004209172446280718, + 0.0032568068709224463, + 0.003980421926826239, + 0.002240363974124193, + 0.0030447018798440695, + 0.0059891087003052235, + 0.0046212319284677505, + 0.006462890654802322, + 0.004403515253216028, + 0.0050234440714120865, + 0.013412454165518284, + 0.004670067690312862, + 0.004663478117436171, + 0.005080068949609995, + 0.006983209867030382, + 0.013822968117892742, + 0.009344827383756638, + 0.016160055994987488, + 0.011261808685958385, + 0.00595256220549345, + 0.013128302060067654, + 0.01209185179322958, + 0.014541249722242355, + 0.05119888857007027, + 0.008488891646265984, + 0.011035625822842121, + 0.017438456416130066, + 0.012256192974746227, + 0.007559021934866905, + 0.008835533633828163, + 0.0065844119526445866, + 0.012903312221169472, + 0.009232211858034134, + 0.01459891814738512, + 0.020528197288513184, + 0.014921063557267189, + 0.012510600499808788, + 0.011353216134011745, + 0.015562871471047401, + 0.04032821208238602, + 0.015312300063669682, + 0.025041108950972557, + 0.029423270374536514, + 0.022097256034612656, + 0.021194929257035255, + 0.016615958884358406, + 0.010193774476647377, + 0.007425250485539436, + 0.004828697070479393, + 0.006198688875883818, + 0.014810572378337383, + 0.0231435839086771, + 0.012840083800256252, + 0.01761721819639206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.005945615470409393 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive watermelon with dark green and pale lemon green stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the watermelon have two thick vine legs emerging from its base that are moving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the watermelon scampering or running across a vast, sunny meadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera positioned at a low angle, close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clear blue sky with a distant, soft-focus treeline?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003191023599356413 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.003191023599356413, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003191023599356413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0785854309797287, + 0.0705249160528183, + 0.02773883379995823, + 0.02906305156648159, + 0.06473773717880249, + 0.09443849325180054, + 0.06953485310077667, + 0.043736446648836136, + 0.22671650350093842, + 0.29492372274398804 + ], + "credit_map": [ + 0.0077971662394702435, + 0.008310560137033463, + 0.009038976393640041, + 0.0110875703394413, + 0.016744114458560944, + 0.00720616988837719, + 0.008831240236759186, + 0.009569641202688217, + 0.005749749485403299, + 0.005588813219219446, + 0.01149697881191969, + 0.005446127615869045, + 0.013551206327974796, + 0.007930811494588852, + 0.008588138036429882, + 0.012173091061413288, + 0.001962882000952959, + 0.0023602060973644257, + 0.0035811173729598522, + 0.00332663394510746, + 0.002176232635974884, + 0.00220403796993196, + 0.007072276435792446, + 0.005055448040366173, + 0.002043464221060276, + 0.002455748151987791, + 0.004328916314989328, + 0.003223516745492816, + 0.003351669292896986, + 0.004024822730571032, + 0.003070715581998229, + 0.0065641989931464195, + 0.006108648609369993, + 0.005845873616635799, + 0.005830969661474228, + 0.003790788585320115, + 0.0210028775036335, + 0.006110805552452803, + 0.004543825052678585, + 0.011503944173455238, + 0.007141195237636566, + 0.008337738923728466, + 0.016293715685606003, + 0.006963081657886505, + 0.008363066241145134, + 0.014882668852806091, + 0.009176169522106647, + 0.023280858993530273, + 0.007419355679303408, + 0.025335822254419327, + 0.003932917024940252, + 0.006511643063277006, + 0.00948324240744114, + 0.007314561866223812, + 0.005188112612813711, + 0.004349199123680592, + 0.0030590335372835398, + 0.0028053473215550184, + 0.015130114741623402, + 0.007210328709334135, + 0.0050185443833470345, + 0.0027563662733882666, + 0.0031630422454327345, + 0.004593667108565569, + 0.024226190522313118, + 0.03546769171953201, + 0.013005401007831097, + 0.022836949676275253, + 0.06291693449020386, + 0.032779280096292496, + 0.02001129277050495, + 0.015472760424017906, + 0.041333042085170746, + 0.021337145939469337, + 0.015628064051270485, + 0.017616303637623787, + 0.04308999329805374, + 0.10173353552818298, + 0.01472692470997572, + 0.0394587442278862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.003191023599356413 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_15/index_tensor([2528, 2528])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_15/index_tensor([2528, 2528])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..02c96a3b186eaf439319a24ba802d3837e1924a6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_15/index_tensor([2528, 2528])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "smoothly rotating, wide-angle shot gracefully sweeps across a world composed entirely of soft pastel strokes, reminiscent of a child's sketch. At the center of the image is a vast, spiraling cloud, representing a mesocyclone. Composed of soft hues of lavender, mint green, and butterscotch yellow, it swirls across the sky with a majestic, slow rhythm. From this vast, flat base, a fantastical rainstorm descends. Countless oversized, teardrop-shaped raindrops in candy colors—including lemon yellow, bubblegum pink, and sky blue—fall vertically in a steady, tranquil rhythm, showering a landscape outlined by simple black lines, featuring stick-figure houses with gabled roofs and lollipop-shaped trees. The entire scene is bathed in an even, shadowless, soft light, like the evenly illuminated pages of a fairytale book, devoid of any distinct light source. The camera's relentless rotation perfectly captures the full scope of this whimsical storm, without a single jerky moment, blending the grandeur of the meteorological phenomenon with the naiveté of the artistic style, creating a dreamlike atmosphere that is both magnificent and profoundly tranquil.", + "seed": 4647, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_15/index_2528_seed_4647_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07373636085540056, + "vlm_logprob": -0.07373636085540056, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.07373636085540056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0062970323488116264 + ], + "opposite_logprob_tokens": [ + -5.2562971115112305 + ], + "logprob_reward": -0.0062970323488116264, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0062970323488116264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3321925699710846, + 0.04322979971766472, + 0.15574368834495544, + 0.05301068723201752, + 0.10171034187078476, + 0.05592327192425728, + 0.023941729217767715, + 0.09415976703166962, + 0.11105640232563019, + 0.029031800106167793 + ], + "credit_map": [ + 0.0165699552744627, + 0.10016311705112457, + 0.01755024865269661, + 0.01818128488957882, + 0.031953729689121246, + 0.10673244297504425, + 0.018217284232378006, + 0.022824522107839584, + 0.00444819126278162, + 0.007372607011348009, + 0.004892390221357346, + 0.006458692718297243, + 0.0072632539086043835, + 0.005453873425722122, + 0.0035068646539002657, + 0.003833926049992442, + 0.01241406798362732, + 0.04964141547679901, + 0.004331650678068399, + 0.006206961814314127, + 0.007618403993546963, + 0.06438654661178589, + 0.006372429430484772, + 0.004772220738232136, + 0.009145505726337433, + 0.008406019769608974, + 0.005271225236356258, + 0.006015730556100607, + 0.008334937505424023, + 0.005205873399972916, + 0.006120314355939627, + 0.004511080216616392, + 0.013403110206127167, + 0.02714955247938633, + 0.003983627539128065, + 0.0037027085199952126, + 0.007189110387116671, + 0.03774755448102951, + 0.004603673238307238, + 0.003931011073291302, + 0.006857441738247871, + 0.006053145043551922, + 0.009506960399448872, + 0.008140540681779385, + 0.005107301287353039, + 0.007516759913414717, + 0.005125988274812698, + 0.007615131791681051, + 0.0027808742597699165, + 0.003784019034355879, + 0.0028481273911893368, + 0.002465955214574933, + 0.0023962545674294233, + 0.004153338726609945, + 0.0016961429500952363, + 0.0038170169573277235, + 0.004640670958906412, + 0.03408990800380707, + 0.0030932219233363867, + 0.004889141768217087, + 0.007145879324525595, + 0.033626165241003036, + 0.003277058247476816, + 0.003397728083655238, + 0.017982978373765945, + 0.013358335942029953, + 0.013427413068711758, + 0.011058966629207134, + 0.01905745081603527, + 0.011256139725446701, + 0.012812897562980652, + 0.012102213688194752, + 0.003550991415977478, + 0.0035717186983674765, + 0.0027809191960841417, + 0.0037412699311971664, + 0.005364678334444761, + 0.004982559476047754, + 0.0027759610675275326, + 0.0022637012880295515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0062970323488116264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195182360708714 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014195182360708714, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195182360708714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4737756848335266, + 0.07059156149625778, + 0.0688931867480278, + 0.04901747405529022, + 0.03709110617637634, + 0.03612368181347847, + 0.06316164880990982, + 0.07750888913869858, + 0.09562636911869049, + 0.028210340067744255 + ], + "credit_map": [ + 0.019590819254517555, + 0.1738288700580597, + 0.014557689428329468, + 0.02059558406472206, + 0.03081478737294674, + 0.17875520884990692, + 0.019358789548277855, + 0.016273999586701393, + 0.008569365367293358, + 0.015482222661376, + 0.007976142689585686, + 0.009850837290287018, + 0.00829267967492342, + 0.00832864549010992, + 0.005195813253521919, + 0.006895860191434622, + 0.009432553313672543, + 0.013567240908741951, + 0.004613954573869705, + 0.003320421325042844, + 0.005778214428573847, + 0.022835351526737213, + 0.005368209909647703, + 0.003977246582508087, + 0.005792580544948578, + 0.008477648720145226, + 0.004691003356128931, + 0.007001938298344612, + 0.008594442158937454, + 0.004138730000704527, + 0.005950557999312878, + 0.004370576236397028, + 0.005542438942939043, + 0.006347505375742912, + 0.0029795465525239706, + 0.0024970965459942818, + 0.004934268072247505, + 0.01018128078430891, + 0.002356726210564375, + 0.0022522483486682177, + 0.004669663496315479, + 0.0041002859361469746, + 0.006033843848854303, + 0.004862752743065357, + 0.004051643889397383, + 0.004847290925681591, + 0.003233138006180525, + 0.004325067158788443, + 0.005653123836964369, + 0.012245816178619862, + 0.010802393779158592, + 0.005418686196208, + 0.004808943718671799, + 0.008664528839290142, + 0.003999747335910797, + 0.011568422429263592, + 0.00403829338029027, + 0.027227628976106644, + 0.0034766558092087507, + 0.004918977618217468, + 0.006261206232011318, + 0.025322481989860535, + 0.002881482010707259, + 0.0033821698743849993, + 0.016946980729699135, + 0.013010975904762745, + 0.011349335312843323, + 0.010977991856634617, + 0.017162982374429703, + 0.009528038091957569, + 0.010028432123363018, + 0.006621645763516426, + 0.0034168811980634928, + 0.003804617328569293, + 0.00279031228274107, + 0.002886179368942976, + 0.004705625120550394, + 0.005561203230172396, + 0.0023817766923457384, + 0.0026637495029717684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.014195182360708714 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059270069003105 + ], + "opposite_logprob_tokens": [ + -3.89559268951416 + ], + "logprob_reward": -0.02059270069003105, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059270069003105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32136911153793335, + 0.049002356827259064, + 0.1664646714925766, + 0.23393400013446808, + 0.023062726482748985, + 0.022747203707695007, + 0.0391935296356678, + 0.02160211093723774, + 0.08847170323133469, + 0.0341525599360466 + ], + "credit_map": [ + 0.015206888318061829, + 0.1063753291964531, + 0.012840552255511284, + 0.014700927771627903, + 0.026765327900648117, + 0.10772471874952316, + 0.01583375595510006, + 0.02192162536084652, + 0.004870159085839987, + 0.008618563413619995, + 0.005621226504445076, + 0.008384844288229942, + 0.00658414838835597, + 0.005632737185806036, + 0.004202591720968485, + 0.005088084377348423, + 0.012793665751814842, + 0.05532046779990196, + 0.005451270379126072, + 0.006441169884055853, + 0.008183447644114494, + 0.06643413007259369, + 0.005959164816886187, + 0.0058813560754060745, + 0.03138023987412453, + 0.041158728301525116, + 0.02449745498597622, + 0.028333287686109543, + 0.040543120354413986, + 0.015337962657213211, + 0.03275830298662186, + 0.019924908876419067, + 0.004369989037513733, + 0.0030241678468883038, + 0.0025564117822796106, + 0.0016896625747904181, + 0.002589197363704443, + 0.004928722511976957, + 0.0019516098545864224, + 0.0019529660930857062, + 0.002896142192184925, + 0.0025431551039218903, + 0.0029370326083153486, + 0.003185100620612502, + 0.0024811250623315573, + 0.0028852836694568396, + 0.0021031901706010103, + 0.003716174280270934, + 0.004228745587170124, + 0.006070094183087349, + 0.006969084497541189, + 0.003719103755429387, + 0.0031868889927864075, + 0.0032582327257841825, + 0.003022201592102647, + 0.00873917993158102, + 0.003587639657780528, + 0.00412251241505146, + 0.001729167765006423, + 0.002108805812895298, + 0.0023680157028138638, + 0.0034256333019584417, + 0.0017360454658046365, + 0.002524293027818203, + 0.010763131082057953, + 0.010838250629603863, + 0.01225723884999752, + 0.011191890574991703, + 0.015517172403633595, + 0.007099370937794447, + 0.01246498804539442, + 0.008339660242199898, + 0.003848402062430978, + 0.007808386813849211, + 0.002475457964465022, + 0.0034890794195234776, + 0.005643962416797876, + 0.006631248164921999, + 0.0023292270489037037, + 0.0019267996540293097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.02059270069003105 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192597009241581 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192597009241581, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192597009241581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3202943503856659, + 0.06429781019687653, + 0.03600465878844261, + 0.05600101873278618, + 0.08313952386379242, + 0.027648430317640305, + 0.053097497671842575, + 0.16059093177318573, + 0.16165094077587128, + 0.03727487474679947 + ], + "credit_map": [ + 0.01794448308646679, + 0.09264064580202103, + 0.017293650656938553, + 0.017949124798178673, + 0.030601250007748604, + 0.09860210865736008, + 0.01819058135151863, + 0.0270724855363369, + 0.006405895575881004, + 0.009898238815367222, + 0.007608265150338411, + 0.011324122548103333, + 0.00996902771294117, + 0.008280256763100624, + 0.005474376957863569, + 0.005337624344974756, + 0.003597608068957925, + 0.006079148035496473, + 0.004468054976314306, + 0.0034592372830957174, + 0.0027328026480972767, + 0.006477368529886007, + 0.00634206086397171, + 0.002848378848284483, + 0.005933948326855898, + 0.007163540925830603, + 0.006387246772646904, + 0.00675039691850543, + 0.007166924886405468, + 0.005066260229796171, + 0.007049812935292721, + 0.010482884012162685, + 0.008114470168948174, + 0.02276325598359108, + 0.006618325598537922, + 0.005051199346780777, + 0.005585447885096073, + 0.02446228452026844, + 0.004023056942969561, + 0.006521478295326233, + 0.0034339085686951876, + 0.002978928852826357, + 0.0036682607606053352, + 0.0035954106133431196, + 0.002861997112631798, + 0.0034373311791568995, + 0.0025157444179058075, + 0.0051568495109677315, + 0.003780339378863573, + 0.011682530865073204, + 0.008092670701444149, + 0.0042348685674369335, + 0.0035741280298680067, + 0.00943778920918703, + 0.003261224366724491, + 0.009033946320414543, + 0.009060300886631012, + 0.05762709677219391, + 0.005987843032926321, + 0.00835366640239954, + 0.011563430540263653, + 0.056820083409547806, + 0.005604990292340517, + 0.005573528818786144, + 0.02291766367852688, + 0.019298195838928223, + 0.023997856304049492, + 0.021825144067406654, + 0.03177157789468765, + 0.012668551877140999, + 0.019508657976984978, + 0.00966329500079155, + 0.004185301251709461, + 0.004525724798440933, + 0.0040010735392570496, + 0.005440425127744675, + 0.0052242944948375225, + 0.005649892147630453, + 0.004325962625443935, + 0.003922199364751577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.014192597009241581 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31340429186820984 + ], + "opposite_logprob_tokens": [ + -1.3134043216705322 + ], + "logprob_reward": -0.31340429186820984, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31340429186820984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24932318925857544, + 0.08578337728977203, + 0.13751870393753052, + 0.05903942137956619, + 0.07869046926498413, + 0.07775253802537918, + 0.08967262506484985, + 0.06402205675840378, + 0.10739562660455704, + 0.050802066922187805 + ], + "credit_map": [ + 0.01973874866962433, + 0.059111498296260834, + 0.019295357167720795, + 0.0322030708193779, + 0.028374046087265015, + 0.04806225746870041, + 0.0212714783847332, + 0.021266696974635124, + 0.009469657205045223, + 0.014784322120249271, + 0.012882277369499207, + 0.010007061064243317, + 0.014163591898977757, + 0.009956841357052326, + 0.007445679511874914, + 0.0070739383809268475, + 0.010035253129899502, + 0.043375466018915176, + 0.007041307631880045, + 0.007543070707470179, + 0.007255570497363806, + 0.04953121766448021, + 0.006785366218537092, + 0.005951434839516878, + 0.006878080777823925, + 0.00987681932747364, + 0.006622712593525648, + 0.008394431322813034, + 0.00918559543788433, + 0.005874133203178644, + 0.007294284645467997, + 0.004913354758173227, + 0.011835725046694279, + 0.011707974597811699, + 0.005762755870819092, + 0.0052221654914319515, + 0.010961601510643959, + 0.019177787005901337, + 0.007038889452815056, + 0.006983560509979725, + 0.011317800730466843, + 0.00873060617595911, + 0.015272552147507668, + 0.00851993914693594, + 0.00918655190616846, + 0.009753590449690819, + 0.007336593698710203, + 0.007634892128407955, + 0.0073821209371089935, + 0.02343655377626419, + 0.009387161582708359, + 0.004732494242489338, + 0.006296808365732431, + 0.02482663094997406, + 0.004035059362649918, + 0.009575791656970978, + 0.004290542099624872, + 0.01892673596739769, + 0.00431154714897275, + 0.006814464461058378, + 0.0053736367262899876, + 0.01671029068529606, + 0.0035742635373026133, + 0.004020569380372763, + 0.02042081579566002, + 0.015163899399340153, + 0.012599175795912743, + 0.014615041203796864, + 0.017617058008909225, + 0.007691345177590847, + 0.010665236040949821, + 0.008623052388429642, + 0.004806610755622387, + 0.006926488596946001, + 0.004837884102016687, + 0.009581259451806545, + 0.006225466262549162, + 0.005632500629872084, + 0.007231769151985645, + 0.005560084246098995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.31340429186820984 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smoothly rotating, wide-angle shot gracefully sweeps across a world composed entirely of soft pastel strokes, reminiscent of a child's sketch. At the center of the image is a vast, spiraling cloud, representing a mesocyclone. Composed of soft hues of lavender, mint green, and butterscotch yellow, it swirls across the sky with a majestic, slow rhythm. From this vast, flat base, a fantastical rainstorm descends. Countless oversized, teardrop-shaped raindrops in candy colors—including lemon yellow, bubblegum pink, and sky blue—fall vertically in a steady, tranquil rhythm, showering a landscape outlined by simple black lines, featuring stick-figure houses with gabled roofs and lollipop-shaped trees. The entire scene is bathed in an even, shadowless, soft light, like the evenly illuminated pages of a fairytale book, devoid of any distinct light source. The camera's relentless rotation perfectly captures the full scope of this whimsical storm, without a single jerky moment, blending the grandeur of the meteorological phenomenon with the naiveté of the artistic style, creating a dreamlike atmosphere that is both magnificent and profoundly tranquil.", + "seed": 4648, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_15/index_2528_seed_4648_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03183362213894725, + "vlm_logprob": -0.03183362213894725, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.03183362213894725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0038580286782234907 + ], + "opposite_logprob_tokens": [ + -6.8788580894470215 + ], + "logprob_reward": -0.0038580286782234907, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0038580286782234907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1948602944612503, + 0.0995742678642273, + 0.04938960447907448, + 0.05903729423880577, + 0.049735840409994125, + 0.07427898794412613, + 0.05316990241408348, + 0.07393492013216019, + 0.18819382786750793, + 0.15782500803470612 + ], + "credit_map": [ + 0.01742963120341301, + 0.021093381568789482, + 0.016634665429592133, + 0.014011147432029247, + 0.0420614629983902, + 0.04510447382926941, + 0.013048644177615643, + 0.02547689527273178, + 0.012029926292598248, + 0.015756288543343544, + 0.014926852658390999, + 0.01698753796517849, + 0.012786362320184708, + 0.008459847420454025, + 0.008449832908809185, + 0.010177616961300373, + 0.005916782189160585, + 0.005585923325270414, + 0.006587406620383263, + 0.00774756446480751, + 0.006660629063844681, + 0.004748587962239981, + 0.00545470230281353, + 0.006688008084893227, + 0.010953588411211967, + 0.005582486744970083, + 0.012563377618789673, + 0.006730998866260052, + 0.005045853555202484, + 0.00507008982822299, + 0.004205127712339163, + 0.008885771967470646, + 0.005315063986927271, + 0.006304580718278885, + 0.008295648731291294, + 0.006089698523283005, + 0.006736266892403364, + 0.005013090558350086, + 0.004069894552230835, + 0.007911593653261662, + 0.010767333209514618, + 0.007908370345830917, + 0.013103364035487175, + 0.006962594110518694, + 0.00826522707939148, + 0.007274177856743336, + 0.006994727998971939, + 0.01300320029258728, + 0.007658952381461859, + 0.005433183163404465, + 0.008966612629592419, + 0.005074417684227228, + 0.005603925324976444, + 0.00745063042268157, + 0.005155959166586399, + 0.007826225832104683, + 0.008702593855559826, + 0.010818022303283215, + 0.01375657320022583, + 0.006616627331823111, + 0.008220584131777287, + 0.007964842021465302, + 0.0062143998220562935, + 0.011641287244856358, + 0.01529762428253889, + 0.04396701231598854, + 0.02198622189462185, + 0.019959306344389915, + 0.03372153639793396, + 0.017830129712820053, + 0.01627296581864357, + 0.019159026443958282, + 0.019324516877532005, + 0.01779697649180889, + 0.019765788689255714, + 0.028320545330643654, + 0.018525363877415657, + 0.018066637217998505, + 0.013336886651813984, + 0.022688303142786026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.0038580286782234907 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076287128031253815 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.0076287128031253815, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076287128031253815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20248766243457794, + 0.07132677733898163, + 0.07196685671806335, + 0.028598599135875702, + 0.033249448984861374, + 0.0320056714117527, + 0.05745173990726471, + 0.026343122124671936, + 0.39813947677612305, + 0.07843069732189178 + ], + "credit_map": [ + 0.014095624908804893, + 0.01952730119228363, + 0.023063182830810547, + 0.02070283517241478, + 0.04725433140993118, + 0.0413963682949543, + 0.014298959635198116, + 0.02214905060827732, + 0.00682039512321353, + 0.008871768601238728, + 0.011830952018499374, + 0.009747256524860859, + 0.010261520743370056, + 0.0068413736298680305, + 0.00756254093721509, + 0.009390974417328835, + 0.00867225881665945, + 0.008128469809889793, + 0.007772581186145544, + 0.008818757720291615, + 0.013664393685758114, + 0.0105719193816185, + 0.006947181653231382, + 0.007391284219920635, + 0.00314453081227839, + 0.0034850609954446554, + 0.005265471525490284, + 0.003555176081135869, + 0.0032271689269691706, + 0.0028488694224506617, + 0.002658970421180129, + 0.004413351882249117, + 0.0032364989165216684, + 0.0034225720446556807, + 0.007162937428802252, + 0.004544271621853113, + 0.0033484732266515493, + 0.003689753357321024, + 0.0031460232567042112, + 0.004698918201029301, + 0.0035477655474096537, + 0.003094029612839222, + 0.0070143225602805614, + 0.0032049622386693954, + 0.0032248503994196653, + 0.003075110260397196, + 0.0029911308083683252, + 0.005853503942489624, + 0.005532338749617338, + 0.005543397273868322, + 0.007432103622704744, + 0.006191306747496128, + 0.009042111225426197, + 0.013755427673459053, + 0.006210475694388151, + 0.003744575660675764, + 0.0030466890893876553, + 0.0034405961632728577, + 0.0036920069251209497, + 0.0026335115544497967, + 0.005351518280804157, + 0.0022008507512509823, + 0.0024799625389277935, + 0.003497987985610962, + 0.045750658959150314, + 0.04957380145788193, + 0.06049461290240288, + 0.040821004658937454, + 0.031001782044768333, + 0.068966805934906, + 0.03690280020236969, + 0.0646279975771904, + 0.008406275883316994, + 0.008686509914696217, + 0.010152308270335197, + 0.015077667310833931, + 0.008431355468928814, + 0.0073606520891189575, + 0.007880537770688534, + 0.012435395270586014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0076287128031253815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188592046499252 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.018188592046499252, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188592046499252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2175317108631134, + 0.04098186641931534, + 0.09242168813943863, + 0.056362878531217575, + 0.05023686960339546, + 0.06331129372119904, + 0.05010310560464859, + 0.048775527626276016, + 0.31721585988998413, + 0.0630592405796051 + ], + "credit_map": [ + 0.021273402497172356, + 0.022411908954381943, + 0.019114118069410324, + 0.02715694159269333, + 0.04964691400527954, + 0.037457648664712906, + 0.018650561571121216, + 0.021820208057761192, + 0.004089300520718098, + 0.0060134935192763805, + 0.005409375764429569, + 0.005840777885168791, + 0.006136625539511442, + 0.004595836624503136, + 0.0043020909652113914, + 0.004594364203512669, + 0.00974461529403925, + 0.009639337658882141, + 0.016800491139292717, + 0.024229273200035095, + 0.0063008652068674564, + 0.008871258236467838, + 0.007140101864933968, + 0.009695738554000854, + 0.007755381055176258, + 0.006399674341082573, + 0.009361065924167633, + 0.0074675725772976875, + 0.007620261516422033, + 0.005544223356992006, + 0.004514350555837154, + 0.007700350135564804, + 0.005647359415888786, + 0.004841307643800974, + 0.011004608124494553, + 0.0063125756569206715, + 0.00451748026534915, + 0.00523404823616147, + 0.005335953086614609, + 0.007343538571149111, + 0.006188796833157539, + 0.00702306954190135, + 0.01150592416524887, + 0.009101717732846737, + 0.006629480980336666, + 0.006952348630875349, + 0.005111264996230602, + 0.010798686183989048, + 0.005869647953659296, + 0.005401360336691141, + 0.008469422347843647, + 0.005266154184937477, + 0.007354378700256348, + 0.008569694124162197, + 0.004512957762926817, + 0.004659492056816816, + 0.00588810071349144, + 0.0065233586356043816, + 0.008220335468649864, + 0.0051991697400808334, + 0.005052547436207533, + 0.006348887924104929, + 0.004395670257508755, + 0.007147456053644419, + 0.03673390671610832, + 0.039865508675575256, + 0.045545052736997604, + 0.033649079501628876, + 0.02630819007754326, + 0.05470309033989906, + 0.03063187561929226, + 0.04977915436029434, + 0.007446134462952614, + 0.00700154947116971, + 0.007780602667480707, + 0.010443967767059803, + 0.007574251387268305, + 0.006525340490043163, + 0.005914508830755949, + 0.010372890159487724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.018188592046499252 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024876620154827833 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024876620154827833, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024876620154827833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.174752339720726, + 0.05450718477368355, + 0.040466900914907455, + 0.030867312103509903, + 0.24105282127857208, + 0.16360889375209808, + 0.06797275692224503, + 0.054665032774209976, + 0.10688978433609009, + 0.06521707028150558 + ], + "credit_map": [ + 0.016756849363446236, + 0.015629692003130913, + 0.023561663925647736, + 0.017815712839365005, + 0.033010248094797134, + 0.030913012102246284, + 0.01480468176305294, + 0.022260461002588272, + 0.008758019655942917, + 0.006226277444511652, + 0.006230240222066641, + 0.007234000600874424, + 0.007508565671741962, + 0.005849089939147234, + 0.00547069264575839, + 0.007230293471366167, + 0.006368611007928848, + 0.003678017994388938, + 0.007381762843579054, + 0.003793147625401616, + 0.0042757438495755196, + 0.005132222082465887, + 0.0032016930636018515, + 0.006635700352489948, + 0.004447206389158964, + 0.002739876974374056, + 0.005356709007173777, + 0.004106271546334028, + 0.0042442213743925095, + 0.0026825405657291412, + 0.003196862991899252, + 0.004093619994819164, + 0.026077184826135635, + 0.01748027466237545, + 0.05267322435975075, + 0.021702641621232033, + 0.018383508548140526, + 0.027675731107592583, + 0.02560071088373661, + 0.051459506154060364, + 0.021074380725622177, + 0.017641371116042137, + 0.02532593533396721, + 0.031320203095674515, + 0.013154975138604641, + 0.015705980360507965, + 0.01344590075314045, + 0.025940118357539177, + 0.01005079410970211, + 0.008332783356308937, + 0.007046947255730629, + 0.00594176584854722, + 0.011568974703550339, + 0.013323402963578701, + 0.007047660648822784, + 0.004660424776375294, + 0.008447249419987202, + 0.005550962872803211, + 0.0076739490032196045, + 0.004700139630585909, + 0.0061964918859303, + 0.010187037289142609, + 0.004757988732308149, + 0.007151207886636257, + 0.012856325134634972, + 0.018835660070180893, + 0.011143983341753483, + 0.013639925979077816, + 0.01565394178032875, + 0.013780626468360424, + 0.011831328272819519, + 0.009147979319095612, + 0.009102998301386833, + 0.007774434983730316, + 0.00878723245114088, + 0.007623602170497179, + 0.006988964043557644, + 0.00924449972808361, + 0.005397684872150421, + 0.010297643952071667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0024876620154827833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12700511515140533 + ], + "opposite_logprob_tokens": [ + -2.127005100250244 + ], + "logprob_reward": -0.12700511515140533, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12700511515140533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08740036189556122, + 0.048750873655080795, + 0.05089025944471359, + 0.04308938980102539, + 0.12728579342365265, + 0.06546700745820999, + 0.09922734647989273, + 0.10895649343729019, + 0.2709819972515106, + 0.09795049577951431 + ], + "credit_map": [ + 0.010541990399360657, + 0.010585155338048935, + 0.008806852623820305, + 0.011502527631819248, + 0.01285616960376501, + 0.013643755577504635, + 0.009332559071481228, + 0.010131356306374073, + 0.006786762271076441, + 0.005396789405494928, + 0.007578723132610321, + 0.006167800631374121, + 0.006192360073328018, + 0.005217837635427713, + 0.005448672454804182, + 0.005961929447948933, + 0.007523218169808388, + 0.005579714197665453, + 0.008527147583663464, + 0.0055178143084049225, + 0.008600065484642982, + 0.005008280742913485, + 0.005106537137180567, + 0.005027481820434332, + 0.006985335610806942, + 0.004268844611942768, + 0.00795390922576189, + 0.005182763561606407, + 0.004826805554330349, + 0.004355179145932198, + 0.0035571034532040358, + 0.0059594507329165936, + 0.017569227144122124, + 0.015425252728164196, + 0.020664246752858162, + 0.015397601760923862, + 0.0147806815803051, + 0.015974199399352074, + 0.011158101260662079, + 0.016316473484039307, + 0.0072814421728253365, + 0.005624521989375353, + 0.012710334733128548, + 0.010305004194378853, + 0.006185915321111679, + 0.007556035649031401, + 0.004711272194981575, + 0.011092482134699821, + 0.008830839768052101, + 0.009414044208824635, + 0.014401351101696491, + 0.00982507411390543, + 0.016285711899399757, + 0.025708671659231186, + 0.009187585674226284, + 0.005574068985879421, + 0.010566899552941322, + 0.020566150546073914, + 0.014116763137280941, + 0.012671470642089844, + 0.015523075126111507, + 0.010687245056033134, + 0.010195482522249222, + 0.014629407785832882, + 0.03060987778007984, + 0.033860962837934494, + 0.04082626476883888, + 0.02856098860502243, + 0.023209281265735626, + 0.045786868780851364, + 0.025480134412646294, + 0.04264762997627258, + 0.009840462356805801, + 0.011725365184247494, + 0.01179470680654049, + 0.02024419605731964, + 0.015258817002177238, + 0.008472132496535778, + 0.007397014647722244, + 0.013217809610068798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.12700511515140533 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smoothly rotating, wide-angle shot gracefully sweeps across a world composed entirely of soft pastel strokes, reminiscent of a child's sketch. At the center of the image is a vast, spiraling cloud, representing a mesocyclone. Composed of soft hues of lavender, mint green, and butterscotch yellow, it swirls across the sky with a majestic, slow rhythm. From this vast, flat base, a fantastical rainstorm descends. Countless oversized, teardrop-shaped raindrops in candy colors—including lemon yellow, bubblegum pink, and sky blue—fall vertically in a steady, tranquil rhythm, showering a landscape outlined by simple black lines, featuring stick-figure houses with gabled roofs and lollipop-shaped trees. The entire scene is bathed in an even, shadowless, soft light, like the evenly illuminated pages of a fairytale book, devoid of any distinct light source. The camera's relentless rotation perfectly captures the full scope of this whimsical storm, without a single jerky moment, blending the grandeur of the meteorological phenomenon with the naiveté of the artistic style, creating a dreamlike atmosphere that is both magnificent and profoundly tranquil.", + "seed": 4649, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_15/index_2528_seed_4649_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09187270496040582, + "vlm_logprob": -0.09187270496040582, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.09187270496040582, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005137693602591753 + ], + "opposite_logprob_tokens": [ + -5.755137920379639 + ], + "logprob_reward": -0.005137693602591753, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.005137693602591753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05016389489173889, + 0.03544335067272186, + 0.02674480713903904, + 0.10248564183712006, + 0.022817961871623993, + 0.025141410529613495, + 0.6725918054580688, + 0.01851567067205906, + 0.03241398558020592, + 0.013681515119969845 + ], + "credit_map": [ + 0.008411814458668232, + 0.009644672274589539, + 0.004197783302515745, + 0.005163002293556929, + 0.009164953604340553, + 0.006156827788800001, + 0.0038004431407898664, + 0.003624402219429612, + 0.005974465515464544, + 0.004251875914633274, + 0.0022828225046396255, + 0.005762490909546614, + 0.004206331912428141, + 0.00441625015810132, + 0.005993484985083342, + 0.0025556283071637154, + 0.004642175510525703, + 0.0030605802312493324, + 0.0014053320046514273, + 0.0024528573267161846, + 0.0044717672280967236, + 0.0071263620629906654, + 0.001924209063872695, + 0.001661523012444377, + 0.01303134299814701, + 0.015112808905541897, + 0.010531282983720303, + 0.010301921516656876, + 0.013341207057237625, + 0.014162994921207428, + 0.011773200705647469, + 0.014230890199542046, + 0.00396316172555089, + 0.003778057638555765, + 0.001922019524499774, + 0.0023514891508966684, + 0.0031864542979747057, + 0.003755368757992983, + 0.0020037093199789524, + 0.0018577012233436108, + 0.00299121649004519, + 0.005019431002438068, + 0.0022753961384296417, + 0.0025005489587783813, + 0.004086317028850317, + 0.00318744289688766, + 0.002558021806180477, + 0.0025230369064956903, + 0.08546837419271469, + 0.08174343407154083, + 0.10744807869195938, + 0.0549011267721653, + 0.07433634251356125, + 0.07301580160856247, + 0.06757453829050064, + 0.1281040757894516, + 0.002223401330411434, + 0.0018108687363564968, + 0.0011814093450084329, + 0.004463608842343092, + 0.0013850133400410414, + 0.0036090509966015816, + 0.0025408421643078327, + 0.0013014745200052857, + 0.004353304393589497, + 0.003450915217399597, + 0.004702535457909107, + 0.0028283291030675173, + 0.0028396504931151867, + 0.004668594337999821, + 0.003368782578036189, + 0.006201873067766428, + 0.0016380802262574434, + 0.0012844084994867444, + 0.001326168654486537, + 0.0021598879247903824, + 0.0012027157936245203, + 0.0015483602182939649, + 0.0026397425681352615, + 0.0018821518169716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.005137693602591753 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187422305345535 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187422305345535, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187422305345535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0906524732708931, + 0.06333038955926895, + 0.14607369899749756, + 0.19640977680683136, + 0.03242919221520424, + 0.024966590106487274, + 0.3432045578956604, + 0.03809923678636551, + 0.05312792956829071, + 0.011706113815307617 + ], + "credit_map": [ + 0.017664464190602303, + 0.012740708887577057, + 0.010706981644034386, + 0.009277892298996449, + 0.010203827172517776, + 0.010098096914589405, + 0.008556196466088295, + 0.011404309421777725, + 0.011570318602025509, + 0.009593449532985687, + 0.004357149824500084, + 0.011934967711567879, + 0.00790287647396326, + 0.007328936830163002, + 0.006408832035958767, + 0.004233855288475752, + 0.025235505774617195, + 0.018525857478380203, + 0.009649952873587608, + 0.017280537635087967, + 0.023330073803663254, + 0.031148040667176247, + 0.010639742948114872, + 0.010263996198773384, + 0.02792966552078724, + 0.031484346836805344, + 0.019194183871150017, + 0.01437582727521658, + 0.03383069485425949, + 0.03159166872501373, + 0.0157710500061512, + 0.02223234251141548, + 0.006263333838433027, + 0.004577265121042728, + 0.002785752061754465, + 0.0044743758626282215, + 0.004200704861432314, + 0.004583202302455902, + 0.002784094540402293, + 0.002760461997240782, + 0.003502767765894532, + 0.003888108767569065, + 0.0032184808515012264, + 0.0027036513201892376, + 0.003898848546668887, + 0.0023971712216734886, + 0.0019586083944886923, + 0.003398953005671501, + 0.04091334715485573, + 0.04196685180068016, + 0.05572673678398132, + 0.02904372662305832, + 0.039259906858205795, + 0.040619149804115295, + 0.03224118426442146, + 0.06343366205692291, + 0.005996516440063715, + 0.0037945006042718887, + 0.0025253561325371265, + 0.0073220026679337025, + 0.00393490819260478, + 0.007463218178600073, + 0.004530197475105524, + 0.0025325347669422626, + 0.007158241234719753, + 0.006167918909341097, + 0.008738254196941853, + 0.004588443785905838, + 0.005218081641942263, + 0.006481990683823824, + 0.00465990649536252, + 0.010115091688930988, + 0.002005473244935274, + 0.00150074134580791, + 0.001243242179043591, + 0.0013890640111640096, + 0.0014401011867448688, + 0.0012747214641422033, + 0.0016492853173986077, + 0.0012034851824864745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.018187422305345535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043065913021564484 + ], + "opposite_logprob_tokens": [ + -3.1680660247802734 + ], + "logprob_reward": -0.043065913021564484, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043065913021564484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07745445519685745, + 0.06274987012147903, + 0.08888953924179077, + 0.5040149092674255, + 0.029482673853635788, + 0.02093745581805706, + 0.13031692802906036, + 0.034570492804050446, + 0.03817969188094139, + 0.013403983786702156 + ], + "credit_map": [ + 0.011966750957071781, + 0.00956545677036047, + 0.011091308668255806, + 0.010830606333911419, + 0.006648554466664791, + 0.008671009913086891, + 0.007880032993853092, + 0.010800727643072605, + 0.00584283284842968, + 0.003590445267036557, + 0.006721922196447849, + 0.01618124544620514, + 0.003765341592952609, + 0.007369257975369692, + 0.014286644756793976, + 0.004992172122001648, + 0.014758951961994171, + 0.0075240228325128555, + 0.008029780350625515, + 0.012811659835278988, + 0.011617830954492092, + 0.019855650141835213, + 0.00757173215970397, + 0.006719901692122221, + 0.07404747605323792, + 0.07057764381170273, + 0.045189280062913895, + 0.029033318161964417, + 0.07395842671394348, + 0.10948169231414795, + 0.044138938188552856, + 0.0575881153345108, + 0.0035285940393805504, + 0.0020766269881278276, + 0.005225612781941891, + 0.0052028982900083065, + 0.001801285776309669, + 0.002796894172206521, + 0.003318657400086522, + 0.005532101262360811, + 0.003084716619923711, + 0.0030830451287329197, + 0.00270821969024837, + 0.0021108132787048817, + 0.002755233319476247, + 0.0024926422629505396, + 0.00229061720892787, + 0.0024121657479554415, + 0.017017588019371033, + 0.018750978633761406, + 0.017892688512802124, + 0.012279564514756203, + 0.016890540719032288, + 0.019565431401133537, + 0.010338827036321163, + 0.01758129894733429, + 0.005324682220816612, + 0.0032436128240078688, + 0.003400718793272972, + 0.006900869309902191, + 0.0033644859213382006, + 0.005139188375324011, + 0.004158704541623592, + 0.003038231050595641, + 0.004877117928117514, + 0.008243496529757977, + 0.003187960246577859, + 0.0030019308906048536, + 0.005487534683197737, + 0.005021947436034679, + 0.0032503921538591385, + 0.00510930921882391, + 0.001662376569584012, + 0.001586079248227179, + 0.0018487347988411784, + 0.0016910930862650275, + 0.0015809126198291779, + 0.001621474977582693, + 0.0013450327096506953, + 0.0020682793110609055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.043065913021564484 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944904405623674 + ], + "opposite_logprob_tokens": [ + -5.130944728851318 + ], + "logprob_reward": -0.005944904405623674, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944904405623674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06771194189786911, + 0.04686259850859642, + 0.15987080335617065, + 0.1488645225763321, + 0.034165505319833755, + 0.03203551471233368, + 0.37750744819641113, + 0.07205117493867874, + 0.038231223821640015, + 0.022699160501360893 + ], + "credit_map": [ + 0.00871970597654581, + 0.012200080789625645, + 0.0063888742588460445, + 0.00694157462567091, + 0.011349675245583057, + 0.008021119982004166, + 0.006588642485439777, + 0.007502275053411722, + 0.0048061516135931015, + 0.005446930881589651, + 0.004743353929370642, + 0.008432911708950996, + 0.0050337244756519794, + 0.005171420052647591, + 0.007687096018344164, + 0.005541018210351467, + 0.027002617716789246, + 0.017836090177297592, + 0.013194712810218334, + 0.018983980640769005, + 0.025307413190603256, + 0.03227177634835243, + 0.012350302189588547, + 0.01292392797768116, + 0.02202197164297104, + 0.020653551444411278, + 0.013444830663502216, + 0.010089448653161526, + 0.02419528178870678, + 0.034255459904670715, + 0.010879558511078358, + 0.013324445113539696, + 0.0043854922987520695, + 0.003832818241789937, + 0.005217001307755709, + 0.0053352732211351395, + 0.0030873219948261976, + 0.003974594175815582, + 0.003105185693129897, + 0.005227824207395315, + 0.003911859821528196, + 0.003473899094387889, + 0.007087595760822296, + 0.0032095005735754967, + 0.0024993228726089, + 0.0035995079670101404, + 0.003407485084608197, + 0.004846346098929644, + 0.0505911149084568, + 0.04784507676959038, + 0.05660375952720642, + 0.030413353815674782, + 0.039866361767053604, + 0.04052165895700455, + 0.0386870875954628, + 0.07297907024621964, + 0.010773205198347569, + 0.006963099353015423, + 0.0069947452284395695, + 0.015262412838637829, + 0.006452891509979963, + 0.011133713647723198, + 0.007985490374267101, + 0.006485625170171261, + 0.005671817343682051, + 0.009426373988389969, + 0.0019248087191954255, + 0.0036248071119189262, + 0.006297689396888018, + 0.0045523555018007755, + 0.00361173995770514, + 0.003121636575087905, + 0.003276263363659382, + 0.00377892074175179, + 0.00258592888712883, + 0.0023499964736402035, + 0.0036774498876184225, + 0.0019795554690063, + 0.002238744404166937, + 0.0028123040683567524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.005944904405623674 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870275914669037 + ], + "opposite_logprob_tokens": [ + -1.137027621269226 + ], + "logprob_reward": -0.3870275914669037, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870275914669037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0234465803951025, + 0.010437308810651302, + 0.025623267516493797, + 0.26230424642562866, + 0.009241041727364063, + 0.010457957163453102, + 0.6213893294334412, + 0.007431509904563427, + 0.018556559458374977, + 0.011112174019217491 + ], + "credit_map": [ + 0.0043137348257005215, + 0.004636646714061499, + 0.001672716811299324, + 0.00250071519985795, + 0.0035156619269400835, + 0.002632227959111333, + 0.0025740410201251507, + 0.001600834890268743, + 0.0016872487030923367, + 0.0017383405938744545, + 0.0008411625167354941, + 0.00140627461951226, + 0.0012511429376900196, + 0.0011309345718473196, + 0.0014572538202628493, + 0.0009249510476365685, + 0.004015205893665552, + 0.003163360757753253, + 0.002122896956279874, + 0.0020213094539940357, + 0.004678500350564718, + 0.005696211010217667, + 0.0017651630332693458, + 0.0021606197115033865, + 0.037123892456293106, + 0.03846073895692825, + 0.02576068602502346, + 0.015754355117678642, + 0.03974195942282677, + 0.04878362640738487, + 0.023304661735892296, + 0.03337432071566582, + 0.002095606876537204, + 0.001102089649066329, + 0.0008296555606648326, + 0.0012840363197028637, + 0.0009281424572691321, + 0.001036645146086812, + 0.0010514406021684408, + 0.0009134243591688573, + 0.0017107477178797126, + 0.001098952954635024, + 0.0011255043791607022, + 0.0015775891952216625, + 0.0010437659220770001, + 0.0016275111120194197, + 0.0011187385534867644, + 0.0011551475618034601, + 0.06941599398851395, + 0.08850224316120148, + 0.09305877238512039, + 0.04958172142505646, + 0.07685364037752151, + 0.07536430656909943, + 0.05738241598010063, + 0.1112302616238594, + 0.0012804099824279547, + 0.0006065957713872194, + 0.0005335621535778046, + 0.0010346835479140282, + 0.0009072180837392807, + 0.0016576439375057817, + 0.0008621218730695546, + 0.0005492744967341423, + 0.003081719856709242, + 0.0039024371653795242, + 0.0017303762724623084, + 0.0017305373912677169, + 0.002916177501901984, + 0.00198321882635355, + 0.0016736261313781142, + 0.0015384666621685028, + 0.0014798498013988137, + 0.0011569088092073798, + 0.0019558779895305634, + 0.0012710121227428317, + 0.0012147607048973441, + 0.001070273108780384, + 0.000916308956220746, + 0.0020471815951168537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.3870275914669037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smoothly rotating, wide-angle shot gracefully sweeps across a world composed entirely of soft pastel strokes, reminiscent of a child's sketch. At the center of the image is a vast, spiraling cloud, representing a mesocyclone. Composed of soft hues of lavender, mint green, and butterscotch yellow, it swirls across the sky with a majestic, slow rhythm. From this vast, flat base, a fantastical rainstorm descends. Countless oversized, teardrop-shaped raindrops in candy colors—including lemon yellow, bubblegum pink, and sky blue—fall vertically in a steady, tranquil rhythm, showering a landscape outlined by simple black lines, featuring stick-figure houses with gabled roofs and lollipop-shaped trees. The entire scene is bathed in an even, shadowless, soft light, like the evenly illuminated pages of a fairytale book, devoid of any distinct light source. The camera's relentless rotation perfectly captures the full scope of this whimsical storm, without a single jerky moment, blending the grandeur of the meteorological phenomenon with the naiveté of the artistic style, creating a dreamlike atmosphere that is both magnificent and profoundly tranquil.", + "seed": 4650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_15/index_2528_seed_4650_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05068714488297701, + "vlm_logprob": -0.05068714488297701, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.05068714488297701, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014056370593607426 + ], + "opposite_logprob_tokens": [ + -4.389056205749512 + ], + "logprob_reward": -0.014056370593607426, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014056370593607426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21159107983112335, + 0.0963236540555954, + 0.0552554577589035, + 0.06559700518846512, + 0.05531668663024902, + 0.03640652075409889, + 0.03139350190758705, + 0.0541551411151886, + 0.12841695547103882, + 0.26554396748542786 + ], + "credit_map": [ + 0.019712384790182114, + 0.025228336453437805, + 0.018754465505480766, + 0.01422025728970766, + 0.05893158167600632, + 0.0411120243370533, + 0.019206007942557335, + 0.014426043257117271, + 0.014303094707429409, + 0.015226046554744244, + 0.00972041953355074, + 0.007974930107593536, + 0.014346536248922348, + 0.015608858317136765, + 0.009271901100873947, + 0.009871873073279858, + 0.007819204591214657, + 0.00801672600209713, + 0.006468332838267088, + 0.004867068491876125, + 0.009992425329983234, + 0.006153843831270933, + 0.005985355004668236, + 0.005952501203864813, + 0.008024604991078377, + 0.007693194784224033, + 0.007673573214560747, + 0.008399236015975475, + 0.0072844610549509525, + 0.009574471041560173, + 0.008239357732236385, + 0.008708100765943527, + 0.005400513764470816, + 0.00965691264718771, + 0.00912934634834528, + 0.005625748075544834, + 0.006051924545317888, + 0.008378027938306332, + 0.005522561725229025, + 0.0055516501888632774, + 0.0041733174584805965, + 0.005552268587052822, + 0.0035658269189298153, + 0.003937178757041693, + 0.00489334762096405, + 0.004918607417494059, + 0.00423957547172904, + 0.005126396659761667, + 0.004631535615772009, + 0.004489003214985132, + 0.00375802512280643, + 0.003034812631085515, + 0.00348176178522408, + 0.004029897041618824, + 0.0035859139170497656, + 0.00438255351036787, + 0.005438788793981075, + 0.008059059269726276, + 0.005166220013052225, + 0.004184093326330185, + 0.01264440082013607, + 0.010561361908912659, + 0.00415485305711627, + 0.003946361597627401, + 0.01535356417298317, + 0.018048647791147232, + 0.013879300095140934, + 0.009003040380775928, + 0.021509427577257156, + 0.015460829250514507, + 0.013802890665829182, + 0.021359259262681007, + 0.035851914435625076, + 0.033019617199897766, + 0.03940713405609131, + 0.018277905881404877, + 0.03087473474442959, + 0.0326920710504055, + 0.029497774317860603, + 0.04592280462384224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.014056370593607426 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184728920459747 + ], + "opposite_logprob_tokens": [ + -4.018184661865234 + ], + "logprob_reward": -0.018184728920459747, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184728920459747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18442021310329437, + 0.08388391137123108, + 0.06984680145978928, + 0.09817173331975937, + 0.06315258145332336, + 0.028737090528011322, + 0.025775877758860588, + 0.040340181440114975, + 0.1433863490819931, + 0.2622852623462677 + ], + "credit_map": [ + 0.014685907401144505, + 0.02818763256072998, + 0.015201818197965622, + 0.012194101698696613, + 0.05371277406811714, + 0.033028826117515564, + 0.014485792256891727, + 0.012923351489007473, + 0.011880861595273018, + 0.01433592289686203, + 0.004199707880616188, + 0.007452988065779209, + 0.01718190312385559, + 0.014601913280785084, + 0.009186032228171825, + 0.0050445725210011005, + 0.009526568464934826, + 0.011835921555757523, + 0.006150359753519297, + 0.006689155474305153, + 0.012051686644554138, + 0.00859496183693409, + 0.009001447819173336, + 0.005996692925691605, + 0.01318870298564434, + 0.014355117455124855, + 0.009402437135577202, + 0.010223436169326305, + 0.01716533489525318, + 0.011163027957081795, + 0.01186872087419033, + 0.01080494374036789, + 0.00590819725766778, + 0.011846678331494331, + 0.008466123603284359, + 0.004901653155684471, + 0.007680699694901705, + 0.006938766222447157, + 0.009776012040674686, + 0.007634447421878576, + 0.0030519035644829273, + 0.0047879330813884735, + 0.0037306444719433784, + 0.0038947679568082094, + 0.004256501793861389, + 0.0033905697055161, + 0.0030823734123259783, + 0.002542395843192935, + 0.0033880313858389854, + 0.004230832681059837, + 0.0025461316108703613, + 0.002827988937497139, + 0.0034068638924509287, + 0.0036300939973443747, + 0.0028898953460156918, + 0.0028560387436300516, + 0.004551379941403866, + 0.009490537457168102, + 0.003448434639722109, + 0.0040781074203550816, + 0.0046630967408418655, + 0.00564993591979146, + 0.00368634145706892, + 0.0047723460011184216, + 0.017895426601171494, + 0.020081624388694763, + 0.0167461596429348, + 0.011025848798453808, + 0.021661991253495216, + 0.019566571339964867, + 0.01563156023621559, + 0.020777162164449692, + 0.03524649515748024, + 0.03423308581113815, + 0.03755318373441696, + 0.018460821360349655, + 0.03072351962327957, + 0.03168044984340668, + 0.030767768621444702, + 0.04361993819475174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.018184728920459747 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003340408205986 + ], + "opposite_logprob_tokens": [ + -2.3503339290618896 + ], + "logprob_reward": -0.1003340408205986, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003340408205986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1641450971364975, + 0.09181693941354752, + 0.10600846260786057, + 0.03912146016955376, + 0.07651994377374649, + 0.03449616581201553, + 0.05597430840134621, + 0.04240325465798378, + 0.081377774477005, + 0.308136522769928 + ], + "credit_map": [ + 0.01796569675207138, + 0.019033532589673996, + 0.017949765548110008, + 0.013640440069139004, + 0.0348176546394825, + 0.02643469162285328, + 0.017722100019454956, + 0.016581229865550995, + 0.013179930858314037, + 0.01352600660175085, + 0.006334612611681223, + 0.0077737849205732346, + 0.014931157231330872, + 0.015984084457159042, + 0.012609779834747314, + 0.007477591745555401, + 0.01410126592963934, + 0.018704719841480255, + 0.009385718032717705, + 0.0103071890771389, + 0.01872061751782894, + 0.012829366140067577, + 0.01251270156353712, + 0.009446884505450726, + 0.006558702792972326, + 0.005227892193943262, + 0.00553388986736536, + 0.003997826483100653, + 0.004603015258908272, + 0.005132615100592375, + 0.003588741412386298, + 0.004478780552744865, + 0.0067732068710029125, + 0.01182310376316309, + 0.011004873551428318, + 0.007378340233117342, + 0.008117981255054474, + 0.009229686111211777, + 0.013307117857038975, + 0.008885638788342476, + 0.003935261629521847, + 0.004918347112834454, + 0.005113297142088413, + 0.004157347604632378, + 0.003925588447600603, + 0.003800660138949752, + 0.004091956652700901, + 0.004553712904453278, + 0.006546535063534975, + 0.006089620757848024, + 0.0073517970740795135, + 0.005233993288129568, + 0.008122259750962257, + 0.008709018118679523, + 0.006101582199335098, + 0.007819502614438534, + 0.004299357999116182, + 0.0074304877780377865, + 0.005421356298029423, + 0.004025622736662626, + 0.005333900451660156, + 0.00682089664041996, + 0.004114621784538031, + 0.004957014229148626, + 0.011826086789369583, + 0.011686029843986034, + 0.008261609822511673, + 0.0058434042148292065, + 0.012045374140143394, + 0.011438735760748386, + 0.00874991063028574, + 0.011526630260050297, + 0.042094744741916656, + 0.04197404161095619, + 0.04268794134259224, + 0.021559014916419983, + 0.03704136610031128, + 0.037099771201610565, + 0.03785702958703041, + 0.047822628170251846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.1003340408205986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577870309352875 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577870309352875, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577870309352875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2160811871290207, + 0.08347015082836151, + 0.05785791575908661, + 0.07036082446575165, + 0.04714639484882355, + 0.02681557461619377, + 0.02941575087606907, + 0.05927099660038948, + 0.13454511761665344, + 0.27503615617752075 + ], + "credit_map": [ + 0.01950390636920929, + 0.02390770986676216, + 0.019571613520383835, + 0.014862045645713806, + 0.06428556889295578, + 0.036632999777793884, + 0.021195368841290474, + 0.016121942549943924, + 0.012060867622494698, + 0.012525612488389015, + 0.0075875455513596535, + 0.007278209086507559, + 0.013150296173989773, + 0.01331793051213026, + 0.009214508347213268, + 0.008335174061357975, + 0.006481448654085398, + 0.00824766419827938, + 0.008246621116995811, + 0.0063320002518594265, + 0.007737411186099052, + 0.006112887989729643, + 0.007541014812886715, + 0.007158859167248011, + 0.01199210062623024, + 0.009898734278976917, + 0.008480722084641457, + 0.008828745223581791, + 0.009603181853890419, + 0.007481731940060854, + 0.007618553005158901, + 0.0064570424146950245, + 0.004374672658741474, + 0.0066842432133853436, + 0.007481431122869253, + 0.004486988764256239, + 0.004709562286734581, + 0.006176093593239784, + 0.006861378904432058, + 0.00637202151119709, + 0.0026756867300719023, + 0.004154975060373545, + 0.0036460922565311193, + 0.004698536358773708, + 0.002778146183118224, + 0.003304773475974798, + 0.003199025522917509, + 0.0023583369329571724, + 0.0035232966765761375, + 0.002894935430958867, + 0.0039980667643249035, + 0.0025730349589139223, + 0.004483850672841072, + 0.005010887514799833, + 0.002700801007449627, + 0.00423087365925312, + 0.0063964915461838245, + 0.00845218077301979, + 0.007776731159538031, + 0.00545392744243145, + 0.009639628231525421, + 0.009487379342317581, + 0.004639612045139074, + 0.007425041403621435, + 0.018276164308190346, + 0.01945214718580246, + 0.014185604639351368, + 0.011148528195917606, + 0.020378591492772102, + 0.018293553963303566, + 0.01404656283557415, + 0.018763957545161247, + 0.03809953108429909, + 0.03754068911075592, + 0.03748827800154686, + 0.01945868693292141, + 0.03315240889787674, + 0.03339112922549248, + 0.032885514199733734, + 0.04301989823579788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.020577870309352875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the artistic style composed of soft pastel strokes resembling a child's sketch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a vast, spiraling cloud representing a mesocyclone visible at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless oversized, teardrop-shaped raindrops in candy colors fall vertically in a steady rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the landscape outlined by simple black lines featuring stick-figure houses with gabled roofs and lollipop-shaped trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, wide-angle rotation to capture the full scope of the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1002827137708664 + ], + "opposite_logprob_tokens": [ + -2.350282669067383 + ], + "logprob_reward": -0.1002827137708664, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1002827137708664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1234956905245781, + 0.1255418062210083, + 0.19989749789237976, + 0.08420689404010773, + 0.05945669114589691, + 0.054697368294000626, + 0.019404852762818336, + 0.02965797856450081, + 0.0607764758169651, + 0.24286481738090515 + ], + "credit_map": [ + 0.015265882946550846, + 0.02066151238977909, + 0.016818027943372726, + 0.011871885508298874, + 0.014676193706691265, + 0.01804688386619091, + 0.013459410518407822, + 0.012695873156189919, + 0.017967699095606804, + 0.024121150374412537, + 0.012429945170879364, + 0.009263791143894196, + 0.01482309028506279, + 0.015443094074726105, + 0.009505050256848335, + 0.021987969055771828, + 0.029387207701802254, + 0.0369616337120533, + 0.013943294063210487, + 0.02164536342024803, + 0.0320969894528389, + 0.023828506469726562, + 0.02803637459874153, + 0.013998118229210377, + 0.015001565217971802, + 0.007168819196522236, + 0.00920980703085661, + 0.006698873359709978, + 0.011137421242892742, + 0.01195268053561449, + 0.010725867934525013, + 0.012311850674450397, + 0.006761710159480572, + 0.012799987569451332, + 0.004498990252614021, + 0.006786613259464502, + 0.010207198560237885, + 0.008070515468716621, + 0.0050842491909861565, + 0.005247420631349087, + 0.00831806194037199, + 0.0064773294143378735, + 0.005975011736154556, + 0.006291714031249285, + 0.006164904218167067, + 0.005903210490942001, + 0.008633404970169067, + 0.0069337268359959126, + 0.003486918518319726, + 0.004174402449280024, + 0.0015865260502323508, + 0.0013115005567669868, + 0.00282577658072114, + 0.00281384470872581, + 0.001295720343478024, + 0.0019101613434031606, + 0.0032072633039206266, + 0.006129543762654066, + 0.0031548351980745792, + 0.0027439938858151436, + 0.0034628165885806084, + 0.003453950397670269, + 0.004474276211112738, + 0.003031297819688916, + 0.011233589611947536, + 0.00818899180740118, + 0.004922109190374613, + 0.005436963867396116, + 0.009785082191228867, + 0.011664239689707756, + 0.004860165063291788, + 0.004685327410697937, + 0.036425087600946426, + 0.034037429839372635, + 0.028446778655052185, + 0.017782999202609062, + 0.030604831874370575, + 0.029101168736815453, + 0.03192588686943054, + 0.03454061597585678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.1002827137708664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_2/index_tensor([4181, 4181])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_2/index_tensor([4181, 4181])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f44d53a30c43bd06b0641f9325a2e9af1aac8c2e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_2/index_tensor([4181, 4181])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The hyper-realistic CGI rendering unfolds in medium close-ups, captured with subtle, organic drifts from a handheld camera, enhancing the scene's intimacy and unease. The camera focuses on a young man standing pensive on an old iron bridge at night. His short, spiky blond hair, dyed platinum, is slightly tousled, as if styled only hours ago, with dark roots clearly visible. His face, with its lifelike skin texture, is etched with a distressed expression; his forehead is slightly wrinkled, his lips are set tightly, and his gaze is downward, lost in thought. Wearing a worn black leather jacket over a simple dark gray T-shirt with a dull sheen, he leans slightly forward, his hands clenched on the cold, damp railing. The cool, silvery moonlight, serving as the primary light source, carves the sharp planes of his face and creates deep shadows, underscoring his somber mood. Below him, the same moonlight reflected and shimmered in a mesmerizing dance on the slow-moving dark water, casting faint, undulating patterns of light onto his face, illuminating the deep, vivid sadness in his eyes and creating an atmosphere of deep melancholy and contemplation.", + "seed": 23196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_2/index_4181_seed_23196_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.032578177750110626, + "vlm_logprob": -0.032578177750110626, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.032578177750110626, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11775092780590057 + ], + "opposite_logprob_tokens": [ + -2.367750883102417 + ], + "logprob_reward": -0.11775092780590057, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11775092780590057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15483780205249786, + 0.06330613791942596, + 0.034122154116630554, + 0.019166437909007072, + 0.031184332445263863, + 0.31866562366485596, + 0.05547076091170311, + 0.06516208499670029, + 0.17795629799365997, + 0.08012834191322327 + ], + "credit_map": [ + 0.025621527805924416, + 0.015683460980653763, + 0.01993585005402565, + 0.026344822719693184, + 0.014344115741550922, + 0.024716924875974655, + 0.014137503691017628, + 0.014053598046302795, + 0.00924734864383936, + 0.006464645732194185, + 0.003732122480869293, + 0.010726998560130596, + 0.007589854300022125, + 0.014387755654752254, + 0.0048623038455843925, + 0.006295109633356333, + 0.0056360517628490925, + 0.0027444057632237673, + 0.0015443802112713456, + 0.006507752928882837, + 0.005388794932514429, + 0.005075386259704828, + 0.0038209392223507166, + 0.0034044398926198483, + 0.003233689581975341, + 0.0012177614262327552, + 0.0014422209933400154, + 0.0029976200312376022, + 0.00207699672318995, + 0.0018826837185770273, + 0.0025328765623271465, + 0.0037825899198651314, + 0.004621356725692749, + 0.0022896951995790005, + 0.003085239790380001, + 0.0049645560793578625, + 0.005567850545048714, + 0.003271599533036351, + 0.004096560645848513, + 0.003287475323304534, + 0.11498571932315826, + 0.010579913854598999, + 0.007750623393803835, + 0.133358895778656, + 0.013187355361878872, + 0.004763728007674217, + 0.025781895965337753, + 0.008257497102022171, + 0.007836442440748215, + 0.0032854622695595026, + 0.01038405392318964, + 0.009337276220321655, + 0.005523243919014931, + 0.004891120828688145, + 0.010440168902277946, + 0.003772994037717581, + 0.00733930803835392, + 0.0046455226838588715, + 0.009633752517402172, + 0.007268531713634729, + 0.007086151745170355, + 0.00979292206466198, + 0.011847836896777153, + 0.007548064924776554, + 0.0254149679094553, + 0.024489935487508774, + 0.024683965370059013, + 0.02432958595454693, + 0.02541271410882473, + 0.0158392284065485, + 0.023469090461730957, + 0.014316819608211517, + 0.013532078824937344, + 0.007524471264332533, + 0.011673228815197945, + 0.01058906503021717, + 0.007901725359261036, + 0.009566554799675941, + 0.011612000875175, + 0.007729214150458574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.11775092780590057 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012559839524328709 + ], + "opposite_logprob_tokens": [ + -4.38755989074707 + ], + "logprob_reward": -0.012559839524328709, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012559839524328709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2621038556098938, + 0.08454938977956772, + 0.07491537928581238, + 0.07187294960021973, + 0.07931078225374222, + 0.07126858085393906, + 0.03699365258216858, + 0.10534756630659103, + 0.1705368161201477, + 0.043101001530885696 + ], + "credit_map": [ + 0.05474943295121193, + 0.023455308750271797, + 0.017979184165596962, + 0.06787066161632538, + 0.024320511147379875, + 0.03071841225028038, + 0.021979378536343575, + 0.021030979230999947, + 0.015198522247374058, + 0.006839666049927473, + 0.005912626627832651, + 0.022589081898331642, + 0.008866941556334496, + 0.00825128611177206, + 0.010107271373271942, + 0.006783993914723396, + 0.013514561578631401, + 0.005925905890762806, + 0.006041642278432846, + 0.015822511166334152, + 0.010459762066602707, + 0.007246414199471474, + 0.009689734317362309, + 0.006214853376150131, + 0.01620040275156498, + 0.004642582964152098, + 0.005549281369894743, + 0.016451066359877586, + 0.009987534023821354, + 0.006799794267863035, + 0.007765110582113266, + 0.004477177746593952, + 0.011813860386610031, + 0.003702523186802864, + 0.0073900786228477955, + 0.01567048579454422, + 0.00997824314981699, + 0.005638340953737497, + 0.016135040670633316, + 0.008982216008007526, + 0.014347638934850693, + 0.00617783609777689, + 0.009847012348473072, + 0.015391887165606022, + 0.0072541553527116776, + 0.005280650686472654, + 0.006997203454375267, + 0.0059721930883824825, + 0.004113845992833376, + 0.003756341291591525, + 0.0032533397898077965, + 0.0045223901979625225, + 0.006858959328383207, + 0.003988819196820259, + 0.00705724535509944, + 0.003442710265517235, + 0.010328199714422226, + 0.008720511570572853, + 0.015528829768300056, + 0.012866580858826637, + 0.013590767048299313, + 0.006931342650204897, + 0.026520848274230957, + 0.010860483162105083, + 0.02043389342725277, + 0.020173244178295135, + 0.028466226533055305, + 0.021451419219374657, + 0.022162528708577156, + 0.018756626173853874, + 0.026677211746573448, + 0.012415649369359016, + 0.004963505547493696, + 0.004936537705361843, + 0.005812490358948708, + 0.00502941245213151, + 0.005886521190404892, + 0.005951132625341415, + 0.006505753844976425, + 0.004015649203211069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.012559839524328709 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779873304069042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779873304069042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779873304069042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2250516712665558, + 0.03345382213592529, + 0.023907797411084175, + 0.022203925997018814, + 0.027233446016907692, + 0.5311194658279419, + 0.01554026734083891, + 0.026211770251393318, + 0.07056491822004318, + 0.024712979793548584 + ], + "credit_map": [ + 0.047883957624435425, + 0.016860712319612503, + 0.013381053693592548, + 0.059012968093156815, + 0.01933462731540203, + 0.03175017237663269, + 0.020227251574397087, + 0.01660093292593956, + 0.005370284430682659, + 0.003471406875178218, + 0.0028153453022241592, + 0.007525513414293528, + 0.0031979361083358526, + 0.004784802440553904, + 0.0033379585947841406, + 0.0029505707789212465, + 0.003671736689284444, + 0.001948493649251759, + 0.002446721773594618, + 0.006396304816007614, + 0.002275514416396618, + 0.0020192416850477457, + 0.003440398722887039, + 0.0017093861242756248, + 0.0031574892345815897, + 0.0014475181233137846, + 0.0022351837251335382, + 0.005106845870614052, + 0.0019216043874621391, + 0.0025962383951991796, + 0.004502405412495136, + 0.0012366415467113256, + 0.004582522436976433, + 0.0017576981335878372, + 0.004428819753229618, + 0.005631234962493181, + 0.0024429576005786657, + 0.002056726487353444, + 0.003234366187825799, + 0.003099119756370783, + 0.19588086009025574, + 0.016765356063842773, + 0.013011951930820942, + 0.2256496101617813, + 0.016352267935872078, + 0.006540174596011639, + 0.04423278570175171, + 0.012686461210250854, + 0.002540484070777893, + 0.0014422157546505332, + 0.0009510829113423824, + 0.0032360670156776905, + 0.00221359939314425, + 0.0015628571854904294, + 0.0023100972175598145, + 0.0012838643742725253, + 0.0036314898170530796, + 0.002042603911831975, + 0.002882098313421011, + 0.0030397046357393265, + 0.002987016225233674, + 0.002646920969709754, + 0.005493917968124151, + 0.0034880186431109905, + 0.008243760094046593, + 0.009448006749153137, + 0.012781847268342972, + 0.008724488317966461, + 0.008441949263215065, + 0.007270549889653921, + 0.010284114629030228, + 0.005370201077312231, + 0.002700316021218896, + 0.0019339564023539424, + 0.0033387052826583385, + 0.003206237917765975, + 0.003578416770324111, + 0.0033234674483537674, + 0.003443236695602536, + 0.0031886431388556957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.009779873304069042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739504635334015 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739504635334015, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739504635334015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13313867151737213, + 0.051609016954898834, + 0.027828363701701164, + 0.023295242339372635, + 0.038062892854213715, + 0.35333773493766785, + 0.051395900547504425, + 0.10872523486614227, + 0.16071689128875732, + 0.05188997834920883 + ], + "credit_map": [ + 0.026467695832252502, + 0.014313827268779278, + 0.012147835455834866, + 0.03235181048512459, + 0.011287772096693516, + 0.018367072567343712, + 0.008889610879123211, + 0.009313061833381653, + 0.00791450496762991, + 0.005032641813158989, + 0.0029266136698424816, + 0.009540989063680172, + 0.0047165341675281525, + 0.013605477288365364, + 0.003838043659925461, + 0.004034220241010189, + 0.004619828891009092, + 0.002243072958663106, + 0.003330494975671172, + 0.004849878139793873, + 0.0035033852327615023, + 0.0035915945190936327, + 0.0035014073364436626, + 0.002188703278079629, + 0.003608860308304429, + 0.0018454630626365542, + 0.0035635700915008783, + 0.004041439387947321, + 0.0024355114437639713, + 0.0026012638118118048, + 0.0028818377759307623, + 0.0023172975052148104, + 0.0069785951636731625, + 0.0020934385247528553, + 0.005883773323148489, + 0.00822425726801157, + 0.004111456219106913, + 0.0029512601904571056, + 0.005351184401661158, + 0.0024689300917088985, + 0.12705804407596588, + 0.012310323305428028, + 0.009331459179520607, + 0.14646033942699432, + 0.012147175148129463, + 0.006254238076508045, + 0.02923595905303955, + 0.010540243238210678, + 0.00707138329744339, + 0.0030993109103292227, + 0.00962839089334011, + 0.007640073075890541, + 0.004347905516624451, + 0.005482029169797897, + 0.00951981171965599, + 0.004607004579156637, + 0.01184666994959116, + 0.006921833846718073, + 0.01797463744878769, + 0.01392379030585289, + 0.01556800864636898, + 0.011849869042634964, + 0.019446706399321556, + 0.011193724349141121, + 0.01949678361415863, + 0.016325030475854874, + 0.02090596966445446, + 0.02155308797955513, + 0.020530207082629204, + 0.017498010769486427, + 0.030205614864826202, + 0.014202200807631016, + 0.007203827612102032, + 0.0035987435840070248, + 0.011547437869012356, + 0.006812210660427809, + 0.004209923557937145, + 0.005203854758292437, + 0.008551212958991528, + 0.004762773867696524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.006739504635334015 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606074348092079 + ], + "opposite_logprob_tokens": [ + -4.141060829162598 + ], + "logprob_reward": -0.01606074348092079, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606074348092079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2413414716720581, + 0.0634487047791481, + 0.041172225028276443, + 0.0435146689414978, + 0.03866863623261452, + 0.34552857279777527, + 0.06287101656198502, + 0.049953609704971313, + 0.08691461384296417, + 0.026586469262838364 + ], + "credit_map": [ + 0.05424192175269127, + 0.01869061589241028, + 0.01876657083630562, + 0.061897896230220795, + 0.02073494903743267, + 0.03601495549082756, + 0.016516001895070076, + 0.014478562399744987, + 0.011456858366727829, + 0.0055475542321801186, + 0.005041873082518578, + 0.017224065959453583, + 0.005002255085855722, + 0.008016333915293217, + 0.0070940423756837845, + 0.004065720830112696, + 0.00663773762062192, + 0.003177823731675744, + 0.0031818258576095104, + 0.010032198391854763, + 0.004557972773909569, + 0.004747319035232067, + 0.0051337541081011295, + 0.003703594207763672, + 0.009505636058747768, + 0.002526947995647788, + 0.004596235230565071, + 0.00926181860268116, + 0.0043223206885159016, + 0.003572287270799279, + 0.005873124580830336, + 0.0038562959525734186, + 0.008439309895038605, + 0.001594756729900837, + 0.0034211487509310246, + 0.010437040589749813, + 0.003490765579044819, + 0.002246442250907421, + 0.006835137493908405, + 0.002204037504270673, + 0.12359058111906052, + 0.01246659830212593, + 0.010415886528789997, + 0.14147262275218964, + 0.011851692572236061, + 0.007563284132629633, + 0.029324479401111603, + 0.00884340237826109, + 0.007043181452900171, + 0.004259833134710789, + 0.011327717453241348, + 0.008274247869849205, + 0.007546456530690193, + 0.008421097882091999, + 0.010663339868187904, + 0.005335147492587566, + 0.005355615634471178, + 0.0048270742408931255, + 0.0037939040921628475, + 0.005262130871415138, + 0.007418971508741379, + 0.007733249105513096, + 0.010581523180007935, + 0.004981142934411764, + 0.008523846045136452, + 0.011778391897678375, + 0.009960848838090897, + 0.011171317659318447, + 0.013201075606048107, + 0.009481817483901978, + 0.01486500445753336, + 0.007932311855256557, + 0.004268747754395008, + 0.0020465387497097254, + 0.00372603302821517, + 0.0032538773957639933, + 0.0031059931498020887, + 0.003383143339306116, + 0.004358199890702963, + 0.002443936187773943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.01606074348092079 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The hyper-realistic CGI rendering unfolds in medium close-ups, captured with subtle, organic drifts from a handheld camera, enhancing the scene's intimacy and unease. The camera focuses on a young man standing pensive on an old iron bridge at night. His short, spiky blond hair, dyed platinum, is slightly tousled, as if styled only hours ago, with dark roots clearly visible. His face, with its lifelike skin texture, is etched with a distressed expression; his forehead is slightly wrinkled, his lips are set tightly, and his gaze is downward, lost in thought. Wearing a worn black leather jacket over a simple dark gray T-shirt with a dull sheen, he leans slightly forward, his hands clenched on the cold, damp railing. The cool, silvery moonlight, serving as the primary light source, carves the sharp planes of his face and creates deep shadows, underscoring his somber mood. Below him, the same moonlight reflected and shimmered in a mesmerizing dance on the slow-moving dark water, casting faint, undulating patterns of light onto his face, illuminating the deep, vivid sadness in his eyes and creating an atmosphere of deep melancholy and contemplation.", + "seed": 23197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_2/index_4181_seed_23197_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.047881044819951056, + "vlm_logprob": -0.047881044819951056, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.047881044819951056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1817304491996765 + ], + "opposite_logprob_tokens": [ + -1.9317305088043213 + ], + "logprob_reward": -0.1817304491996765, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1817304491996765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2514449656009674, + 0.09809883683919907, + 0.09080471098423004, + 0.2281099557876587, + 0.03275880590081215, + 0.058764342218637466, + 0.060933176428079605, + 0.06060352176427841, + 0.05959915369749069, + 0.0588824599981308 + ], + "credit_map": [ + 0.03717910125851631, + 0.03852755203843117, + 0.018690617755055428, + 0.03861261531710625, + 0.029041731730103493, + 0.032370761036872864, + 0.03200867399573326, + 0.025013919919729233, + 0.01653907261788845, + 0.01601812057197094, + 0.010652170516550541, + 0.013161951676011086, + 0.008464813232421875, + 0.015303854830563068, + 0.010270701721310616, + 0.00768816564232111, + 0.01453584898263216, + 0.010220101103186607, + 0.007674273569136858, + 0.01462713535875082, + 0.008426391519606113, + 0.009223545901477337, + 0.015464197844266891, + 0.010633236728608608, + 0.06633875519037247, + 0.013122113421559334, + 0.0189504511654377, + 0.01633668690919876, + 0.012313180603086948, + 0.016476690769195557, + 0.06466177850961685, + 0.019910311326384544, + 0.006914148107171059, + 0.0029523649718612432, + 0.0017975213704630733, + 0.00392568064853549, + 0.0021086495835334063, + 0.00447107432410121, + 0.007428495679050684, + 0.0031608755234628916, + 0.008603433147072792, + 0.007850799709558487, + 0.0059386626817286015, + 0.007532715331763029, + 0.00501707149669528, + 0.007136746309697628, + 0.010878107510507107, + 0.005806812550872564, + 0.009446397423744202, + 0.007407013792544603, + 0.007479460444301367, + 0.005677582696080208, + 0.0050298720598220825, + 0.0059199281968176365, + 0.007629581727087498, + 0.012343345209956169, + 0.013001023791730404, + 0.004504451993852854, + 0.004498927388340235, + 0.007146285381168127, + 0.0044809021055698395, + 0.00976317934691906, + 0.01039204653352499, + 0.006816714070737362, + 0.009204036556184292, + 0.00678664818406105, + 0.005661551374942064, + 0.006729545071721077, + 0.0066907634027302265, + 0.008353287354111671, + 0.009714527055621147, + 0.006458797957748175, + 0.008724969811737537, + 0.0051834010519087315, + 0.006550709251314402, + 0.008564370684325695, + 0.004999345634132624, + 0.009082059375941753, + 0.007997420616447926, + 0.007780187297612429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.1817304491996765 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01422197837382555 + ], + "opposite_logprob_tokens": [ + -4.264222145080566 + ], + "logprob_reward": -0.01422197837382555, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01422197837382555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.139034703373909, + 0.08643738180398941, + 0.1163402795791626, + 0.3619522750377655, + 0.04156314209103584, + 0.06868532299995422, + 0.04123757407069206, + 0.06917987018823624, + 0.031230458989739418, + 0.04433893412351608 + ], + "credit_map": [ + 0.023490283638238907, + 0.026007387787103653, + 0.0102281104773283, + 0.02027846872806549, + 0.013586660847067833, + 0.015343754552304745, + 0.01886119693517685, + 0.011238844133913517, + 0.013492284342646599, + 0.018869813531637192, + 0.006115424912422895, + 0.013172298669815063, + 0.007495731581002474, + 0.009497864171862602, + 0.009560981765389442, + 0.008232985623180866, + 0.022628147155046463, + 0.019572360441088676, + 0.011690902523696423, + 0.017943277955055237, + 0.012751211412250996, + 0.007422084454447031, + 0.01583702862262726, + 0.008495260961353779, + 0.13233858346939087, + 0.013415951281785965, + 0.026009151712059975, + 0.018889358267188072, + 0.017255229875445366, + 0.015991520136594772, + 0.11880332231521606, + 0.019249169155955315, + 0.010799325071275234, + 0.003404507879167795, + 0.003070332808420062, + 0.004230671562254429, + 0.002165835816413164, + 0.004872934892773628, + 0.009931208565831184, + 0.0030883229337632656, + 0.011037155985832214, + 0.009859498590230942, + 0.008777312934398651, + 0.006705689709633589, + 0.005805624183267355, + 0.009950255043804646, + 0.009437990374863148, + 0.007111794780939817, + 0.006336066871881485, + 0.005297556519508362, + 0.004001364577561617, + 0.004140523727983236, + 0.004946169443428516, + 0.005975157488137484, + 0.0054258364252746105, + 0.005114899482578039, + 0.012975954450666904, + 0.008770755492150784, + 0.005423691123723984, + 0.009788556024432182, + 0.007031798362731934, + 0.007305394858121872, + 0.010051297023892403, + 0.007832425646483898, + 0.005026268307119608, + 0.003920208662748337, + 0.002091215457767248, + 0.005153766367584467, + 0.002122627105563879, + 0.004363475367426872, + 0.005512500647455454, + 0.003040397772565484, + 0.006426586303859949, + 0.006444405764341354, + 0.0037427940405905247, + 0.007227605674415827, + 0.004974353592842817, + 0.004626487381756306, + 0.006491989362984896, + 0.004404712468385696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.01422197837382555 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073797941207886 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073797941207886, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073797941207886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09940196573734283, + 0.06689193099737167, + 0.1144862100481987, + 0.12038499861955643, + 0.1578528881072998, + 0.04827702417969704, + 0.06070270761847496, + 0.15841086208820343, + 0.10891123116016388, + 0.06468022614717484 + ], + "credit_map": [ + 0.016051093116402626, + 0.010034877806901932, + 0.008378596976399422, + 0.021250901743769646, + 0.008093401789665222, + 0.012475972063839436, + 0.015815909951925278, + 0.007301213685423136, + 0.009661905467510223, + 0.010849371552467346, + 0.005199784878641367, + 0.010294823907315731, + 0.0074666752479970455, + 0.010181126184761524, + 0.009555313736200333, + 0.00368293235078454, + 0.017847921699285507, + 0.010038843378424644, + 0.014900671318173409, + 0.017993154004216194, + 0.00617425749078393, + 0.012949206866323948, + 0.025601491332054138, + 0.008980653248727322, + 0.03649306669831276, + 0.006600141525268555, + 0.010697992518544197, + 0.01106349565088749, + 0.006136447656899691, + 0.007029389496892691, + 0.03261791169643402, + 0.009746546857059002, + 0.057828523218631744, + 0.00815046951174736, + 0.006553263403475285, + 0.01017052587121725, + 0.007436560932546854, + 0.005324816331267357, + 0.054378800094127655, + 0.00800990592688322, + 0.010454115457832813, + 0.0055733914487063885, + 0.0051505002193152905, + 0.0048342514783144, + 0.0031281800474971533, + 0.005916585214436054, + 0.00908961147069931, + 0.00413038395345211, + 0.013234901241958141, + 0.005113225430250168, + 0.005182987079024315, + 0.004334183409810066, + 0.004860007669776678, + 0.00983337126672268, + 0.008944200351834297, + 0.00919982697814703, + 0.031233519315719604, + 0.013101032935082912, + 0.014892674051225185, + 0.020744292065501213, + 0.009079879149794579, + 0.024762075394392014, + 0.022385673597455025, + 0.0222117118537426, + 0.01777089573442936, + 0.009418039582669735, + 0.007544644642621279, + 0.014978313818573952, + 0.008108274079859257, + 0.017812086269259453, + 0.020483266562223434, + 0.0127957072108984, + 0.009084959514439106, + 0.005584547761827707, + 0.0068635279312729836, + 0.008049332536756992, + 0.009433805011212826, + 0.0085605438798666, + 0.009508369490504265, + 0.0075951311737298965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.011073797941207886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201411046087742 + ], + "opposite_logprob_tokens": [ + -4.264201641082764 + ], + "logprob_reward": -0.014201411046087742, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201411046087742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14988407492637634, + 0.12436794489622116, + 0.08802998065948486, + 0.46430498361587524, + 0.04911264404654503, + 0.023492593318223953, + 0.021472636610269547, + 0.02523413486778736, + 0.034005485475063324, + 0.020095519721508026 + ], + "credit_map": [ + 0.028248295187950134, + 0.012897319160401821, + 0.019048180431127548, + 0.019654415547847748, + 0.013783710077404976, + 0.024425599724054337, + 0.018176766112446785, + 0.013649782165884972, + 0.02232437953352928, + 0.01345918420702219, + 0.017411764711141586, + 0.014099457301199436, + 0.01119113340973854, + 0.015570039860904217, + 0.017045622691512108, + 0.013266360387206078, + 0.01786518469452858, + 0.008576614782214165, + 0.014754834584891796, + 0.009595202282071114, + 0.0059322151355445385, + 0.01158242579549551, + 0.011781773529946804, + 0.00794172752648592, + 0.17589092254638672, + 0.015729738399386406, + 0.030356112867593765, + 0.02513180673122406, + 0.01977083459496498, + 0.02032676711678505, + 0.15557830035686493, + 0.021520527079701424, + 0.014649019576609135, + 0.0030670787673443556, + 0.003185272216796875, + 0.003918183967471123, + 0.002313164761289954, + 0.005131944082677364, + 0.013013012707233429, + 0.003834968898445368, + 0.003429304575547576, + 0.0033855843357741833, + 0.0024341167882084846, + 0.003003162331879139, + 0.0021663096267729998, + 0.002925330772995949, + 0.003459241474047303, + 0.002689544577151537, + 0.003378533525392413, + 0.002463193843141198, + 0.0020858044736087322, + 0.0019350313814356923, + 0.00204974552616477, + 0.0027545487973839045, + 0.002881921362131834, + 0.003923858515918255, + 0.0046554249711334705, + 0.002886012429371476, + 0.0020540442783385515, + 0.0039029985200613737, + 0.0013652611523866653, + 0.003253679256886244, + 0.004403972532600164, + 0.0027127419598400593, + 0.005374823696911335, + 0.003521219128742814, + 0.003563100239261985, + 0.004915955942124128, + 0.0027681521605700254, + 0.004343007691204548, + 0.005072901025414467, + 0.004446324426680803, + 0.002658572979271412, + 0.001939295558258891, + 0.0021312700118869543, + 0.002971241483464837, + 0.00332642812281847, + 0.002371705137193203, + 0.0023359486367553473, + 0.002361058723181486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.014201411046087742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018177587538957596 + ], + "opposite_logprob_tokens": [ + -4.018177509307861 + ], + "logprob_reward": -0.018177587538957596, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018177587538957596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21109913289546967, + 0.06494320929050446, + 0.06465782225131989, + 0.1655041128396988, + 0.1527961641550064, + 0.05770101770758629, + 0.06728542596101761, + 0.0647493526339531, + 0.08446677029132843, + 0.06679702550172806 + ], + "credit_map": [ + 0.03296737000346184, + 0.02445232681930065, + 0.02524985745549202, + 0.039167068898677826, + 0.02341599389910698, + 0.022897809743881226, + 0.023942377418279648, + 0.019006328657269478, + 0.010499532334506512, + 0.009149516001343727, + 0.006791951600462198, + 0.007517312653362751, + 0.005571983754634857, + 0.009691495448350906, + 0.00833086296916008, + 0.007390553597360849, + 0.014715636149048805, + 0.006857977714389563, + 0.008422289974987507, + 0.008247695863246918, + 0.004759036935865879, + 0.005179672501981258, + 0.012161895632743835, + 0.004313617013394833, + 0.040149345993995667, + 0.013978391885757446, + 0.018159162253141403, + 0.013585628010332584, + 0.01125603448599577, + 0.020913001149892807, + 0.033206429332494736, + 0.014256126247346401, + 0.05558544024825096, + 0.008986732922494411, + 0.005129215773195028, + 0.010455171577632427, + 0.007036601658910513, + 0.006329434458166361, + 0.0534772090613842, + 0.005796364042907953, + 0.008354179561138153, + 0.007451698184013367, + 0.00685251597315073, + 0.006534348241984844, + 0.004979421850293875, + 0.008566298522055149, + 0.008223622106015682, + 0.006738931406289339, + 0.010398009791970253, + 0.008314554579555988, + 0.006540338974446058, + 0.006044927053153515, + 0.004996309522539377, + 0.005297467578202486, + 0.008669913746416569, + 0.017023906111717224, + 0.011053883470594883, + 0.008818094618618488, + 0.006646140944212675, + 0.005849973764270544, + 0.004713214468210936, + 0.009902359917759895, + 0.006248402874916792, + 0.011517281644046307, + 0.012339750304818153, + 0.017092926427721977, + 0.004922015126794577, + 0.0156732015311718, + 0.007834397256374359, + 0.00823700800538063, + 0.011658016592264175, + 0.006709448993206024, + 0.008124974556267262, + 0.007468858268111944, + 0.006976699456572533, + 0.008043131791055202, + 0.012417390942573547, + 0.007393841166049242, + 0.007379522547125816, + 0.00899260863661766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.018177587538957596 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The hyper-realistic CGI rendering unfolds in medium close-ups, captured with subtle, organic drifts from a handheld camera, enhancing the scene's intimacy and unease. The camera focuses on a young man standing pensive on an old iron bridge at night. His short, spiky blond hair, dyed platinum, is slightly tousled, as if styled only hours ago, with dark roots clearly visible. His face, with its lifelike skin texture, is etched with a distressed expression; his forehead is slightly wrinkled, his lips are set tightly, and his gaze is downward, lost in thought. Wearing a worn black leather jacket over a simple dark gray T-shirt with a dull sheen, he leans slightly forward, his hands clenched on the cold, damp railing. The cool, silvery moonlight, serving as the primary light source, carves the sharp planes of his face and creates deep shadows, underscoring his somber mood. Below him, the same moonlight reflected and shimmered in a mesmerizing dance on the slow-moving dark water, casting faint, undulating patterns of light onto his face, illuminating the deep, vivid sadness in his eyes and creating an atmosphere of deep melancholy and contemplation.", + "seed": 23198, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_2/index_4181_seed_23198_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020733061153441666, + "vlm_logprob": -0.020733061153441666, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.020733061153441666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05704711377620697 + ], + "opposite_logprob_tokens": [ + -3.182047128677368 + ], + "logprob_reward": -0.05704711377620697, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05704711377620697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14820438623428345, + 0.2003139853477478, + 0.048958320170640945, + 0.042599666863679886, + 0.03615298867225647, + 0.06647400557994843, + 0.4110724925994873, + 0.019437404349446297, + 0.013759652152657509, + 0.013027064502239227 + ], + "credit_map": [ + 0.025576073676347733, + 0.014856263063848019, + 0.01828368939459324, + 0.015488149598240852, + 0.013587253168225288, + 0.02124924771487713, + 0.020899077877402306, + 0.018264632672071457, + 0.024139506742358208, + 0.009907985106110573, + 0.005861422047019005, + 0.007525320164859295, + 0.010092442855238914, + 0.0767350122332573, + 0.013203432783484459, + 0.052848849445581436, + 0.004711853340268135, + 0.004515366163104773, + 0.002968804445117712, + 0.007027016952633858, + 0.00916249118745327, + 0.006094609387218952, + 0.008286704309284687, + 0.006191475782543421, + 0.007079131901264191, + 0.002466934034600854, + 0.004162456374615431, + 0.005538997706025839, + 0.00470337551087141, + 0.006025900598615408, + 0.006205464247614145, + 0.006417405791580677, + 0.003556897398084402, + 0.004601520486176014, + 0.0019199884263798594, + 0.004925720859318972, + 0.005478984676301479, + 0.005561309400945902, + 0.005256874952465296, + 0.00485169468447566, + 0.0046097454614937305, + 0.005909188650548458, + 0.006088661961257458, + 0.00907503254711628, + 0.007276540622115135, + 0.010251265950500965, + 0.012104327790439129, + 0.011159239336848259, + 0.031183913350105286, + 0.015101839788258076, + 0.012059327214956284, + 0.022615712136030197, + 0.020201822742819786, + 0.029640205204486847, + 0.1520015150308609, + 0.12826816737651825, + 0.0017274044221267104, + 0.0020112621132284403, + 0.002201935276389122, + 0.0018780396785587072, + 0.002257679123431444, + 0.002486195182427764, + 0.004158426076173782, + 0.0027164635248482227, + 0.0012402634602040052, + 0.0010037178872153163, + 0.0010953937890008092, + 0.001974693266674876, + 0.0023215466644614935, + 0.0018219627672806382, + 0.0021630963310599327, + 0.0021389781031757593, + 0.0010126224951818585, + 0.0010077898623421788, + 0.000796781328972429, + 0.002433170098811388, + 0.0013961100485175848, + 0.0024698851630091667, + 0.0022867212537676096, + 0.0016239838441833854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.05704711377620697 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014219275675714016 + ], + "opposite_logprob_tokens": [ + -4.264219284057617 + ], + "logprob_reward": -0.014219275675714016, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014219275675714016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1528874784708023, + 0.17711086571216583, + 0.06742049008607864, + 0.060745321214199066, + 0.054623477160930634, + 0.05759380757808685, + 0.24696972966194153, + 0.13042914867401123, + 0.02677369862794876, + 0.025446034967899323 + ], + "credit_map": [ + 0.02511586621403694, + 0.014295379631221294, + 0.0142526151612401, + 0.01606733351945877, + 0.026102613657712936, + 0.01749439537525177, + 0.019763492047786713, + 0.019795771688222885, + 0.03825440630316734, + 0.013635436072945595, + 0.009135814383625984, + 0.013049147091805935, + 0.012826823629438877, + 0.044596362859010696, + 0.01876996085047722, + 0.026842890307307243, + 0.006499955430626869, + 0.00600447366014123, + 0.0037238062359392643, + 0.009465927258133888, + 0.012098328210413456, + 0.007059187162667513, + 0.013964257203042507, + 0.008604546077549458, + 0.010326582007110119, + 0.0029164976440370083, + 0.004101600032299757, + 0.006379355676472187, + 0.008270672522485256, + 0.007423515897244215, + 0.01091297809034586, + 0.01041411329060793, + 0.005876546259969473, + 0.005458636675029993, + 0.0025464643258601427, + 0.01063165906816721, + 0.009803666733205318, + 0.008688910864293575, + 0.006630517542362213, + 0.004987069871276617, + 0.0055380817502737045, + 0.0054403627291321754, + 0.006939019076526165, + 0.005776640959084034, + 0.003905679564923048, + 0.012951304204761982, + 0.007641238626092672, + 0.009401476010680199, + 0.018404491245746613, + 0.01187430415302515, + 0.0071483636274933815, + 0.011449414305388927, + 0.014172077178955078, + 0.027061328291893005, + 0.08284629881381989, + 0.07401341199874878, + 0.008983194828033447, + 0.01691683754324913, + 0.008977624587714672, + 0.015640070661902428, + 0.02026098221540451, + 0.015753380954265594, + 0.02476229891180992, + 0.01913474127650261, + 0.0017485131975263357, + 0.0021269619464874268, + 0.0020806288812309504, + 0.004273152910172939, + 0.004194813314825296, + 0.002901155734434724, + 0.005363330245018005, + 0.004085139837116003, + 0.0018919316353276372, + 0.0033876863308250904, + 0.001804532716050744, + 0.003156482707709074, + 0.004042043350636959, + 0.004161912016570568, + 0.003635214874520898, + 0.003366228425875306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.014219275675714016 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069775447249413 + ], + "opposite_logprob_tokens": [ + -4.1410698890686035 + ], + "logprob_reward": -0.016069775447249413, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069775447249413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1483418047428131, + 0.3461828827857971, + 0.06390547752380371, + 0.055544402450323105, + 0.09419597685337067, + 0.07069915533065796, + 0.08967141807079315, + 0.0418732576072216, + 0.03523885831236839, + 0.054346825927495956 + ], + "credit_map": [ + 0.027336904779076576, + 0.016485264524817467, + 0.009086180478334427, + 0.019633833318948746, + 0.021254325285553932, + 0.0156424418091774, + 0.020248543471097946, + 0.018654314801096916, + 0.045838672667741776, + 0.0343683660030365, + 0.027464859187602997, + 0.0352482795715332, + 0.03367018699645996, + 0.04528609663248062, + 0.04680785909295082, + 0.07749852538108826, + 0.007220391649752855, + 0.004230601247400045, + 0.006484250072389841, + 0.008317534811794758, + 0.009358011186122894, + 0.007559059653431177, + 0.012091577053070068, + 0.00864405371248722, + 0.011208931915462017, + 0.003200042759999633, + 0.0025912688579410315, + 0.00523215439170599, + 0.007783833425492048, + 0.006722266785800457, + 0.00870540365576744, + 0.010100500658154488, + 0.011904355138540268, + 0.008816556073725224, + 0.00551283173263073, + 0.014550158753991127, + 0.015913790091872215, + 0.015613546594977379, + 0.012124229222536087, + 0.009760511107742786, + 0.006503827404230833, + 0.006435737945139408, + 0.007116954308003187, + 0.008524712175130844, + 0.006950725801289082, + 0.014829202555119991, + 0.010965770110487938, + 0.009372229687869549, + 0.006166389677673578, + 0.005597251933068037, + 0.0026803258806467056, + 0.00558788375928998, + 0.0069306776858866215, + 0.011455371975898743, + 0.028322720900177956, + 0.022930797189474106, + 0.004085129592567682, + 0.004082472529262304, + 0.003915165085345507, + 0.003874811576679349, + 0.0043907007202506065, + 0.007370791397988796, + 0.007689609192311764, + 0.0064645796082913876, + 0.0027291153091937304, + 0.002973473398014903, + 0.0030243797227740288, + 0.005002021323889494, + 0.0057447231374681, + 0.00628322921693325, + 0.004675019532442093, + 0.004806900396943092, + 0.003985318820923567, + 0.005270797293633223, + 0.004353659227490425, + 0.008380880579352379, + 0.007327597588300705, + 0.009507608599960804, + 0.008235001005232334, + 0.007285964209586382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.016069775447249413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052565219812095165 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.0052565219812095165, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052565219812095165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.355398565530777, + 0.13264219462871552, + 0.03457304462790489, + 0.07764798402786255, + 0.05820784717798233, + 0.08916272968053818, + 0.13354358077049255, + 0.05745893344283104, + 0.033859532326459885, + 0.027505487203598022 + ], + "credit_map": [ + 0.07128141820430756, + 0.040644075721502304, + 0.04952160641551018, + 0.037958547472953796, + 0.02795257419347763, + 0.036285292357206345, + 0.0438089482486248, + 0.04794611036777496, + 0.014461199752986431, + 0.007889838889241219, + 0.007229360286146402, + 0.005817359779030085, + 0.008019435219466686, + 0.03975564241409302, + 0.010128203779459, + 0.03934115543961525, + 0.004781875293701887, + 0.0029419127386063337, + 0.0025236180517822504, + 0.004292575642466545, + 0.005276783369481564, + 0.0043986183591187, + 0.006188042927533388, + 0.004169619642198086, + 0.01282514538615942, + 0.004175401292741299, + 0.006609789561480284, + 0.009393147192895412, + 0.009243668988347054, + 0.011768819764256477, + 0.011485363356769085, + 0.012146648950874805, + 0.006508484482765198, + 0.005009364802390337, + 0.003936265595257282, + 0.008917155675590038, + 0.009278428740799427, + 0.010516280308365822, + 0.008453241549432278, + 0.005588624160736799, + 0.007810065057128668, + 0.010129385627806187, + 0.009529460221529007, + 0.012682745233178139, + 0.0076362453401088715, + 0.016470815986394882, + 0.013556844554841518, + 0.011347169987857342, + 0.013990247622132301, + 0.009129545651376247, + 0.010180460289120674, + 0.009913104586303234, + 0.012454301118850708, + 0.02704179286956787, + 0.02766447700560093, + 0.02316965162754059, + 0.005838032346218824, + 0.005256816744804382, + 0.004913920070976019, + 0.00552762858569622, + 0.006282296031713486, + 0.007657856680452824, + 0.013031755574047565, + 0.008950628340244293, + 0.00237392820417881, + 0.0022932353895157576, + 0.003281816141679883, + 0.004368760623037815, + 0.0038081516977399588, + 0.005419603548943996, + 0.0065006958320736885, + 0.0058133406564593315, + 0.002194833941757679, + 0.002236026804894209, + 0.0014258548617362976, + 0.004971500486135483, + 0.003775441087782383, + 0.0034782702568918467, + 0.005213385913521051, + 0.004210175946354866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0052565219812095165 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072618886828423 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072618886828423, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072618886828423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18638934195041656, + 0.26511305570602417, + 0.10239856690168381, + 0.14895832538604736, + 0.03401492163538933, + 0.07608290761709213, + 0.05011380463838577, + 0.05494602397084236, + 0.02655286155641079, + 0.055430177599191666 + ], + "credit_map": [ + 0.042205825448036194, + 0.01932985708117485, + 0.012859106995165348, + 0.019345242530107498, + 0.02505888231098652, + 0.025466229766607285, + 0.018456358462572098, + 0.023667840287089348, + 0.027711335569620132, + 0.024091478437185287, + 0.011228896677494049, + 0.01656455174088478, + 0.02215973660349846, + 0.07838735729455948, + 0.02199763059616089, + 0.06297207623720169, + 0.011337082833051682, + 0.010711969807744026, + 0.011002404615283012, + 0.008528350852429867, + 0.016496235504746437, + 0.013772426173090935, + 0.01668160781264305, + 0.013868489302694798, + 0.03277882933616638, + 0.012624282389879227, + 0.006128810346126556, + 0.013419578783214092, + 0.020550785586237907, + 0.012896466068923473, + 0.022391343489289284, + 0.028168225660920143, + 0.004115642048418522, + 0.004722949583083391, + 0.002218357054516673, + 0.0047025554813444614, + 0.004351210314780474, + 0.005191980395466089, + 0.004849476274102926, + 0.0038627474568784237, + 0.005875805858522654, + 0.014765494503080845, + 0.0061092618852853775, + 0.012482373043894768, + 0.007276963908225298, + 0.00879566092044115, + 0.011423083953559399, + 0.009354259818792343, + 0.004821401089429855, + 0.005447214003652334, + 0.00445153471082449, + 0.0045875427313148975, + 0.005939524620771408, + 0.007119245361536741, + 0.010190458036959171, + 0.007556886877864599, + 0.005527082830667496, + 0.005519711412489414, + 0.004421314690262079, + 0.005732409656047821, + 0.006503203418105841, + 0.007471826858818531, + 0.011365710757672787, + 0.008404762484133244, + 0.002575853606685996, + 0.002159607131034136, + 0.0026625883765518665, + 0.002564093330875039, + 0.0029776685405522585, + 0.005562292877584696, + 0.004070217255502939, + 0.003980540204793215, + 0.004656617995351553, + 0.00650088069960475, + 0.0032422186341136694, + 0.006761013530194759, + 0.009979928843677044, + 0.00719562079757452, + 0.01003499235957861, + 0.007058907765895128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.011072618886828423 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The hyper-realistic CGI rendering unfolds in medium close-ups, captured with subtle, organic drifts from a handheld camera, enhancing the scene's intimacy and unease. The camera focuses on a young man standing pensive on an old iron bridge at night. His short, spiky blond hair, dyed platinum, is slightly tousled, as if styled only hours ago, with dark roots clearly visible. His face, with its lifelike skin texture, is etched with a distressed expression; his forehead is slightly wrinkled, his lips are set tightly, and his gaze is downward, lost in thought. Wearing a worn black leather jacket over a simple dark gray T-shirt with a dull sheen, he leans slightly forward, his hands clenched on the cold, damp railing. The cool, silvery moonlight, serving as the primary light source, carves the sharp planes of his face and creates deep shadows, underscoring his somber mood. Below him, the same moonlight reflected and shimmered in a mesmerizing dance on the slow-moving dark water, casting faint, undulating patterns of light onto his face, illuminating the deep, vivid sadness in his eyes and creating an atmosphere of deep melancholy and contemplation.", + "seed": 23199, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_2/index_4181_seed_23199_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02135273306630552, + "vlm_logprob": -0.02135273306630552, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02135273306630552, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022837484255433083 + ], + "opposite_logprob_tokens": [ + -4.0228376388549805 + ], + "logprob_reward": -0.022837484255433083, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.022837484255433083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1645924150943756, + 0.06726496666669846, + 0.057538386434316635, + 0.036720454692840576, + 0.08073098957538605, + 0.05005528777837753, + 0.055850595235824585, + 0.03894691914319992, + 0.060722049325704575, + 0.38757801055908203 + ], + "credit_map": [ + 0.03047395497560501, + 0.010627065785229206, + 0.03542470559477806, + 0.012385092675685883, + 0.00832001119852066, + 0.040566109120845795, + 0.016423190012574196, + 0.010372273623943329, + 0.008010189048945904, + 0.0054318867623806, + 0.01868182234466076, + 0.005225372035056353, + 0.004132784903049469, + 0.013417411595582962, + 0.008403233252465725, + 0.003962256945669651, + 0.006591977085918188, + 0.004351347219198942, + 0.012037177570164204, + 0.009203426539897919, + 0.003315276699140668, + 0.008913767524063587, + 0.007249310612678528, + 0.005876100156456232, + 0.0049863047897815704, + 0.002187349135056138, + 0.007440104614943266, + 0.00625039916485548, + 0.003406606847420335, + 0.003885115496814251, + 0.005840549245476723, + 0.0027240265626460314, + 0.010739360004663467, + 0.005311018321663141, + 0.013102144934237003, + 0.009701354429125786, + 0.013677735812962055, + 0.007922199554741383, + 0.011473958380520344, + 0.008803206495940685, + 0.007395865861326456, + 0.002833946142345667, + 0.01185766700655222, + 0.004661987069994211, + 0.004747652914375067, + 0.005381657276302576, + 0.006560761481523514, + 0.006615744903683662, + 0.004831835627555847, + 0.003929954022169113, + 0.0069707161746919155, + 0.009027870371937752, + 0.011702334508299828, + 0.0055424910970032215, + 0.008527180179953575, + 0.005318204872310162, + 0.002855980070307851, + 0.003558836877346039, + 0.0035245029721409082, + 0.005564810708165169, + 0.010249834507703781, + 0.004006132483482361, + 0.006109355483204126, + 0.0030774602200835943, + 0.004905676003545523, + 0.006971146911382675, + 0.006149888504296541, + 0.005232425406575203, + 0.014837505295872688, + 0.007887043058872223, + 0.009974864311516285, + 0.004763493780046701, + 0.02078721486032009, + 0.022036893293261528, + 0.01714096963405609, + 0.020764730870723724, + 0.02290576696395874, + 0.12533122301101685, + 0.13835668563842773, + 0.0202544666826725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.022837484255433083 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048662614077329636 + ], + "opposite_logprob_tokens": [ + -3.0486626625061035 + ], + "logprob_reward": -0.048662614077329636, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048662614077329636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060676395893096924, + 0.06413903087377548, + 0.2068193107843399, + 0.2210509181022644, + 0.019588693976402283, + 0.022077856585383415, + 0.0750638097524643, + 0.019770313054323196, + 0.03287619352340698, + 0.27793747186660767 + ], + "credit_map": [ + 0.009298835881054401, + 0.004303019028156996, + 0.009224328212440014, + 0.00740434555336833, + 0.006103625055402517, + 0.007628834340721369, + 0.010744997300207615, + 0.005968407727777958, + 0.008956089615821838, + 0.005324722267687321, + 0.010118948295712471, + 0.008856658823788166, + 0.0057736956514418125, + 0.008747733198106289, + 0.010068883188068867, + 0.006292297039180994, + 0.023173091933131218, + 0.01621687039732933, + 0.03167878836393356, + 0.028858067467808723, + 0.027993708848953247, + 0.026425529271364212, + 0.0329870879650116, + 0.019486170262098312, + 0.026925956830382347, + 0.010071873664855957, + 0.025715218856930733, + 0.03930830582976341, + 0.02134028635919094, + 0.024956271052360535, + 0.05432874336838722, + 0.018404260277748108, + 0.002728492021560669, + 0.0011103239376097918, + 0.0024544496554881334, + 0.0027752178721129894, + 0.0027397910598665476, + 0.002052811672911048, + 0.002679216442629695, + 0.003048389684408903, + 0.0019355211406946182, + 0.0021329079754650593, + 0.0026994284708052874, + 0.0023427281994372606, + 0.00500451447442174, + 0.002345883985981345, + 0.0032609959598630667, + 0.0023558749817311764, + 0.007034599781036377, + 0.0048866597935557365, + 0.012817793525755405, + 0.007652851287275553, + 0.017456624656915665, + 0.007705968338996172, + 0.008804724551737309, + 0.008704591542482376, + 0.0022641366813331842, + 0.0015247141709551215, + 0.0028829481452703476, + 0.002231516409665346, + 0.0038399475160986185, + 0.001726892776787281, + 0.003181760897859931, + 0.0021183977369219065, + 0.004883924033492804, + 0.0032332062255591154, + 0.0027272976003587246, + 0.002816940890625119, + 0.005455144681036472, + 0.005819207523018122, + 0.004772214684635401, + 0.0031682581175118685, + 0.015211285091936588, + 0.014194558374583721, + 0.009601698257029057, + 0.013668186962604523, + 0.01708190143108368, + 0.0965220183134079, + 0.09973368793725967, + 0.011924145743250847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.048662614077329636 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107332669198513 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.01107332669198513, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107332669198513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09404108673334122, + 0.14427003264427185, + 0.05028877779841423, + 0.0568174310028553, + 0.058307792991399765, + 0.04128744453191757, + 0.06460820883512497, + 0.0602436289191246, + 0.09236273169517517, + 0.3377728760242462 + ], + "credit_map": [ + 0.01744714565575123, + 0.005444761365652084, + 0.017927205190062523, + 0.01153646968305111, + 0.006276490166783333, + 0.013954226858913898, + 0.012337458319962025, + 0.00911732204258442, + 0.017693081870675087, + 0.014184893108904362, + 0.02855348028242588, + 0.019832376390695572, + 0.010194949805736542, + 0.016393208876252174, + 0.023106088861823082, + 0.014311926439404488, + 0.005704224109649658, + 0.005070779472589493, + 0.005593412555754185, + 0.007930236868560314, + 0.006684751249849796, + 0.006205953191965818, + 0.007999500259757042, + 0.0050999196246266365, + 0.004499215167015791, + 0.0029114065691828728, + 0.0051065050065517426, + 0.010109211318194866, + 0.007609119638800621, + 0.006951100658625364, + 0.014811472035944462, + 0.004819394554942846, + 0.008275050669908524, + 0.004045384936034679, + 0.008079981431365013, + 0.006119497586041689, + 0.01059260405600071, + 0.005964514799416065, + 0.008803869597613811, + 0.006426887586712837, + 0.004696978256106377, + 0.0036930267233401537, + 0.003227940993383527, + 0.0051062460988759995, + 0.008947959169745445, + 0.0045569464564323425, + 0.006602537352591753, + 0.004455805290490389, + 0.008038467727601528, + 0.0034005572088062763, + 0.0161025058478117, + 0.005305029917508364, + 0.010015815496444702, + 0.0066383229568600655, + 0.007574817165732384, + 0.007532692980021238, + 0.007744044531136751, + 0.003969152458012104, + 0.008588520810008049, + 0.00695448461920023, + 0.009248885326087475, + 0.006757826078683138, + 0.009616339579224586, + 0.0073643699288368225, + 0.008181972429156303, + 0.012893804349005222, + 0.005535545758903027, + 0.009369676001369953, + 0.025832535699009895, + 0.01046051736921072, + 0.014424928463995457, + 0.005663742311298847, + 0.01988283544778824, + 0.01617281697690487, + 0.015482881106436253, + 0.014726322144269943, + 0.020474018529057503, + 0.11671773344278336, + 0.11701633781194687, + 0.017299916595220566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.01107332669198513 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036162238102406263 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.0036162238102406263, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036162238102406263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28417789936065674, + 0.05805318057537079, + 0.04072399064898491, + 0.026514321565628052, + 0.07861220836639404, + 0.019584959372878075, + 0.023440515622496605, + 0.023947417736053467, + 0.01686437986791134, + 0.4280811846256256 + ], + "credit_map": [ + 0.04352415353059769, + 0.014501182362437248, + 0.07735694199800491, + 0.02599945478141308, + 0.012388492003083229, + 0.056767772883176804, + 0.033310726284980774, + 0.020329143851995468, + 0.008017667569220066, + 0.004738008137792349, + 0.013941233977675438, + 0.0051108761690557, + 0.004993811249732971, + 0.0082710525020957, + 0.008769235573709011, + 0.004211292136460543, + 0.005579050164669752, + 0.0037108149845153093, + 0.006485889200121164, + 0.006418789736926556, + 0.0037283499259501696, + 0.005392765160650015, + 0.005619433242827654, + 0.003788893111050129, + 0.0042162309400737286, + 0.0016338148852810264, + 0.004643658641725779, + 0.0035054185427725315, + 0.002476068679243326, + 0.0030154548585414886, + 0.004554667975753546, + 0.002469004364684224, + 0.013111293315887451, + 0.0047278208658099174, + 0.0106428237631917, + 0.011174987070262432, + 0.0070714266039431095, + 0.008020007982850075, + 0.012235120870172977, + 0.011628720909357071, + 0.0028996814507991076, + 0.001573721063323319, + 0.00203651818446815, + 0.0024205476511269808, + 0.004348481073975563, + 0.0018987144576385617, + 0.002832656493410468, + 0.0015746366698294878, + 0.0032934658229351044, + 0.0015576531877741218, + 0.002510352758690715, + 0.0026782448403537273, + 0.005470622796565294, + 0.002232208149507642, + 0.0037816993426531553, + 0.0019162667449563742, + 0.002119456883519888, + 0.0022070787381380796, + 0.0023755687288939953, + 0.002748197643086314, + 0.006205539684742689, + 0.002594070741906762, + 0.0034000768791884184, + 0.00229742843657732, + 0.0014424716355279088, + 0.0018450649222359061, + 0.0015488367062062025, + 0.00211215834133327, + 0.004344948101788759, + 0.001544028869830072, + 0.002672031754627824, + 0.0013548387214541435, + 0.022704962641000748, + 0.023815900087356567, + 0.015728674829006195, + 0.02245298959314823, + 0.024694543331861496, + 0.14141911268234253, + 0.15611377358436584, + 0.02115117572247982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.0036162238102406263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at night on an old iron bridge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character a young man with short, spiky blond hair that is dyed platinum with visible dark roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a worn black leather jacket over a dark gray T-shirt with a dull sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by cool, silvery moonlight that creates deep shadows and reflects on dark water below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium close-up with subtle, organic drifts from a handheld style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020574016496539116 + ], + "opposite_logprob_tokens": [ + -3.8955740928649902 + ], + "logprob_reward": -0.020574016496539116, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020574016496539116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07483343034982681, + 0.06359951943159103, + 0.04384995624423027, + 0.10414627194404602, + 0.05438905209302902, + 0.0217277929186821, + 0.02191341109573841, + 0.043768733739852905, + 0.03261934220790863, + 0.5391525626182556 + ], + "credit_map": [ + 0.011909892782568932, + 0.00523374741896987, + 0.013003801926970482, + 0.007532085292041302, + 0.004673846065998077, + 0.014520719647407532, + 0.011751522310078144, + 0.006207805126905441, + 0.009492143988609314, + 0.006676299497485161, + 0.013744063675403595, + 0.007448886055499315, + 0.005003609228879213, + 0.00873546116054058, + 0.00689129950478673, + 0.005607747007161379, + 0.0064905486069619656, + 0.005898966453969479, + 0.007023077458143234, + 0.005425687879323959, + 0.003355570835992694, + 0.007023932877928019, + 0.005410488229244947, + 0.00322167482227087, + 0.017697589471936226, + 0.00647989846765995, + 0.021763235330581665, + 0.012322353199124336, + 0.009106352925300598, + 0.009430203586816788, + 0.016741948202252388, + 0.010604672133922577, + 0.007177333347499371, + 0.0028610352892428637, + 0.007687568198889494, + 0.0065131173469126225, + 0.008281374350190163, + 0.005535199772566557, + 0.00793609581887722, + 0.008397314697504044, + 0.0017140737036243081, + 0.0020973074715584517, + 0.002433736575767398, + 0.003118432592600584, + 0.003889783751219511, + 0.0020048394799232483, + 0.003532273229211569, + 0.0029373406432569027, + 0.0029595184605568647, + 0.0017881367821246386, + 0.002216609660536051, + 0.0030516034457832575, + 0.004064772743731737, + 0.0022803328465670347, + 0.003219184698536992, + 0.0023332489654421806, + 0.0028899535536766052, + 0.004595572128891945, + 0.002918137004598975, + 0.006172676105052233, + 0.013272009789943695, + 0.004474250599741936, + 0.006261036731302738, + 0.003185086650773883, + 0.004807546269148588, + 0.002255588537082076, + 0.003717999439686537, + 0.002267492702230811, + 0.00400595972314477, + 0.00671664671972394, + 0.004821506794542074, + 0.0040265945717692375, + 0.02789611928164959, + 0.027898689731955528, + 0.016675349324941635, + 0.025514397770166397, + 0.03138513118028641, + 0.1876164674758911, + 0.19867399334907532, + 0.02349233068525791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.020574016496539116 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_3/index_tensor([6281, 6281])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_3/index_tensor([6281, 6281])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3a3b31eedf9ca8ca5b93736a2b8d15b000e710ff --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_3/index_tensor([6281, 6281])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic shot, initially static, captures two humanoid figures composed entirely of delicate, overlapping pink rose petals, engaging in a graceful courtship dance amidst a surreal, dreamlike landscape. Their intricately layered forms, with petals ranging from soft blush to deep burgundy, create a sense of depth and texture, while their movements are fluid and incredibly graceful. They twirl and embrace, their petal-laden limbs entwined; during their embrace, their forms temporarily merge into a beautiful blossom, individual petals swirling together before gently pulling apart, perfectly reforming their individual forms. The ground beneath their feet is soft, luminous moss, bathing the scene in an ethereal, diffuse light from an unseen source, casting no harsh shadows. As their enchanting ritual continues, the camera begins a slow, seamless crane shot, rising vertically to reveal the vastness of their surroundings. The two figures are part of a vast, blooming garden, where countless other flowers – lilies, tulips and peonies – also take on human shapes and perform a similar, mesmerizing dance, creating a silent, synchronized ballet across a dreamlike terrain, fostering a profound atmosphere of tranquility, magic and romantic wonder.", + "seed": 12530, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_3/index_6281_seed_12530_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1437021903693676, + "vlm_logprob": -0.1437021903693676, + "vlm_yes_no_margin": 2.499999964237213, + "vlm_token_credit": -0.1437021903693676, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02405412122607231 + ], + "opposite_logprob_tokens": [ + -4.274054050445557 + ], + "logprob_reward": -0.02405412122607231, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02405412122607231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06020897999405861, + 0.17044322192668915, + 0.40919241309165955, + 0.02040429785847664, + 0.04762387275695801, + 0.09492451697587967, + 0.025596458464860916, + 0.047513287514448166, + 0.05851902440190315, + 0.06557388603687286 + ], + "credit_map": [ + 0.005454717203974724, + 0.006084920838475227, + 0.004762513563036919, + 0.0047011952847242355, + 0.007943956181406975, + 0.004393580835312605, + 0.020727617666125298, + 0.006140480749309063, + 0.04142650216817856, + 0.014718716964125633, + 0.016007449477910995, + 0.01803136058151722, + 0.017651010304689407, + 0.011710026301443577, + 0.020341815426945686, + 0.030556339770555496, + 0.025119485333561897, + 0.01938851736485958, + 0.09391306340694427, + 0.025010252371430397, + 0.14681476354599, + 0.022799266502261162, + 0.05093712359666824, + 0.025209955871105194, + 0.0023773876018822193, + 0.0019526408286765218, + 0.0029948409646749496, + 0.003399956738576293, + 0.0022647420410066843, + 0.0022570344153791666, + 0.002963706385344267, + 0.0021939906291663647, + 0.003084923606365919, + 0.004486693535000086, + 0.007301309145987034, + 0.007568823639303446, + 0.0034257241059094667, + 0.003052839543670416, + 0.010120952501893044, + 0.008582609705626965, + 0.006196389440447092, + 0.004625146742910147, + 0.007930716499686241, + 0.004639161750674248, + 0.008103256113827229, + 0.008361552841961384, + 0.030166080221533775, + 0.024902209639549255, + 0.0027371239848434925, + 0.003184962086379528, + 0.0037499871104955673, + 0.0034140602219849825, + 0.0033797246869653463, + 0.0034828491043299437, + 0.003059327369555831, + 0.002588423201814294, + 0.005206640809774399, + 0.005067397374659777, + 0.010712097398936749, + 0.004947671201080084, + 0.004698829259723425, + 0.0037259443197399378, + 0.009127690456807613, + 0.004027016460895538, + 0.006498751696199179, + 0.008947335183620453, + 0.008395329117774963, + 0.009133304469287395, + 0.005749411880970001, + 0.008061963133513927, + 0.006871191319078207, + 0.004861733876168728, + 0.007788501679897308, + 0.00868536252528429, + 0.006344376131892204, + 0.008063792251050472, + 0.01171500887721777, + 0.01018282026052475, + 0.007751812692731619, + 0.005042212549597025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.02405412122607231 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905363827943802 + ], + "opposite_logprob_tokens": [ + -2.4640536308288574 + ], + "logprob_reward": -0.08905363827943802, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905363827943802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06674981117248535, + 0.1498691290616989, + 0.279843270778656, + 0.05551829934120178, + 0.09661633521318436, + 0.0394480898976326, + 0.040895670652389526, + 0.04758535325527191, + 0.17697444558143616, + 0.04649960249662399 + ], + "credit_map": [ + 0.007837182842195034, + 0.013599754311144352, + 0.00411993870511651, + 0.003311011241748929, + 0.004773005843162537, + 0.005532199051231146, + 0.01919683814048767, + 0.008379879407584667, + 0.03430870547890663, + 0.015488862060010433, + 0.013356059789657593, + 0.01657373458147049, + 0.015849420800805092, + 0.01085707638412714, + 0.019446317106485367, + 0.02398895099759102, + 0.01585615798830986, + 0.013600513339042664, + 0.05253778398036957, + 0.017263038083910942, + 0.10130152106285095, + 0.017742885276675224, + 0.044512417167425156, + 0.017028922215104103, + 0.005059703718870878, + 0.006512111984193325, + 0.0060913898050785065, + 0.012533154338598251, + 0.007313597481697798, + 0.00468880170956254, + 0.00877930223941803, + 0.004540237598121166, + 0.00851733423769474, + 0.008059507235884666, + 0.01099823135882616, + 0.012383867055177689, + 0.006254958920180798, + 0.00715983472764492, + 0.0241093672811985, + 0.019133228808641434, + 0.0025945925153791904, + 0.0016295788809657097, + 0.00345515925437212, + 0.0016479423502460122, + 0.0033036493696272373, + 0.004026446957141161, + 0.01205533929169178, + 0.01073538139462471, + 0.004789018537849188, + 0.0042157587595283985, + 0.005146493203938007, + 0.004390805494040251, + 0.005470392759889364, + 0.006757179740816355, + 0.006138721480965614, + 0.003987297881394625, + 0.0053429435938596725, + 0.0034687549341470003, + 0.011651294305920601, + 0.002449386054649949, + 0.005503728520125151, + 0.0024532447569072247, + 0.013354619033634663, + 0.0033613741397857666, + 0.017782971262931824, + 0.025039587169885635, + 0.019764894619584084, + 0.03183595463633537, + 0.020497653633356094, + 0.025677256286144257, + 0.020098792389035225, + 0.016277335584163666, + 0.005872371140867472, + 0.007785429246723652, + 0.005620637908577919, + 0.005792598705738783, + 0.006967281457036734, + 0.006448278669267893, + 0.004618714097887278, + 0.003394292900338769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.08905363827943802 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370317071676254 + ], + "opposite_logprob_tokens": [ + -3.408703088760376 + ], + "logprob_reward": -0.03370317071676254, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370317071676254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09388791024684906, + 0.037988968193531036, + 0.37835174798965454, + 0.20086972415447235, + 0.035387612879276276, + 0.033789489418268204, + 0.0745275467634201, + 0.08732005208730698, + 0.015559731051325798, + 0.04231724143028259 + ], + "credit_map": [ + 0.009148018434643745, + 0.02185586653649807, + 0.005057589616626501, + 0.004793553613126278, + 0.006485411431640387, + 0.0077988263219594955, + 0.03035608120262623, + 0.008392569608986378, + 0.004208723548799753, + 0.0061374143697321415, + 0.004402256105095148, + 0.004757639952003956, + 0.0077029564417898655, + 0.0026088841259479523, + 0.0043424200266599655, + 0.003828674554824829, + 0.023447681218385696, + 0.01846144162118435, + 0.08190098404884338, + 0.022448301315307617, + 0.1362195611000061, + 0.025215379893779755, + 0.048766277730464935, + 0.021892119199037552, + 0.02785506099462509, + 0.017616644501686096, + 0.024329712614417076, + 0.03447351232171059, + 0.020956886932253838, + 0.021331898868083954, + 0.03721083328127861, + 0.017095183953642845, + 0.0027558926958590746, + 0.0029536238871514797, + 0.0047415210865437984, + 0.005396487191319466, + 0.0023896563798189163, + 0.0018726881826296449, + 0.00817007478326559, + 0.007107670418918133, + 0.0021170140244066715, + 0.001665786956436932, + 0.0032672311645001173, + 0.0014294892316684127, + 0.0024741424713283777, + 0.0025253864005208015, + 0.010347624309360981, + 0.009962817654013634, + 0.008282274007797241, + 0.007801721803843975, + 0.009591873735189438, + 0.009950689971446991, + 0.0076879230327904224, + 0.008844805881381035, + 0.013966255821287632, + 0.008402002975344658, + 0.011504043824970722, + 0.008630883879959583, + 0.0165583286434412, + 0.007848599925637245, + 0.010105200111865997, + 0.007392487023025751, + 0.018373072147369385, + 0.006907448172569275, + 0.0017345944652333856, + 0.0028263903222978115, + 0.0014018648071214557, + 0.0023532125633209944, + 0.0019822916947305202, + 0.00207901862449944, + 0.0020717333536595106, + 0.0011106267338618636, + 0.00499306432902813, + 0.005863910540938377, + 0.004681117832660675, + 0.005697396583855152, + 0.007577926851809025, + 0.005557049531489611, + 0.005039607640355825, + 0.0029071669559925795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.03370317071676254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42888861894607544 + ], + "opposite_logprob_tokens": [ + -1.0538885593414307 + ], + "logprob_reward": -0.42888861894607544, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42888861894607544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07457096874713898, + 0.2486894726753235, + 0.17152968049049377, + 0.046643175184726715, + 0.08368199318647385, + 0.05315285548567772, + 0.11373758316040039, + 0.1602506786584854, + 0.024942800402641296, + 0.022800780832767487 + ], + "credit_map": [ + 0.005749920848757029, + 0.016147861257195473, + 0.004547836724668741, + 0.004208268132060766, + 0.005137131083756685, + 0.004001756198704243, + 0.02786063961684704, + 0.006917556282132864, + 0.06185800954699516, + 0.02076595462858677, + 0.02175184153020382, + 0.024360090494155884, + 0.027606217190623283, + 0.017294570803642273, + 0.03386298567056656, + 0.04118978604674339, + 0.011860838159918785, + 0.008327837102115154, + 0.0410722978413105, + 0.010564576834440231, + 0.06143590807914734, + 0.009573237970471382, + 0.01829337701201439, + 0.010401601903140545, + 0.005962136667221785, + 0.004348881542682648, + 0.0052617439068853855, + 0.008720675483345985, + 0.005546610336750746, + 0.004060235805809498, + 0.008679045364260674, + 0.004063844680786133, + 0.00895223394036293, + 0.009378441609442234, + 0.012125393375754356, + 0.009929610416293144, + 0.008294171653687954, + 0.006828000769019127, + 0.01651071570813656, + 0.011663423851132393, + 0.005338241811841726, + 0.002274787984788418, + 0.00492712901905179, + 0.0020256994757801294, + 0.004952548537403345, + 0.008458572439849377, + 0.013349706307053566, + 0.011826169677078724, + 0.013341519981622696, + 0.01288988534361124, + 0.014384202659130096, + 0.016610821709036827, + 0.013251308351755142, + 0.01295480877161026, + 0.017217624932527542, + 0.013087410479784012, + 0.01992681995034218, + 0.017683658748865128, + 0.02480301260948181, + 0.015279073268175125, + 0.020171187818050385, + 0.012675963342189789, + 0.0346403643488884, + 0.015070594847202301, + 0.003084376687183976, + 0.0037135889288038015, + 0.0045598894357681274, + 0.004659606609493494, + 0.002268279902637005, + 0.0024871628265827894, + 0.0026318058371543884, + 0.0015380908735096455, + 0.003313168417662382, + 0.003287125378847122, + 0.0035506763961166143, + 0.0021606944501399994, + 0.0023862342350184917, + 0.004263470880687237, + 0.0017996946116909385, + 0.002039715414866805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.42888861894607544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281140267848969 + ], + "opposite_logprob_tokens": [ + -2.0178112983703613 + ], + "logprob_reward": -0.14281140267848969, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14281140267848969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23560026288032532, + 0.2953721880912781, + 0.017168479040265083, + 0.01827116496860981, + 0.1173006147146225, + 0.03710918128490448, + 0.06968370079994202, + 0.1630110740661621, + 0.012121226638555527, + 0.0343620590865612 + ], + "credit_map": [ + 0.01739232987165451, + 0.05229654163122177, + 0.011576084420084953, + 0.010942347347736359, + 0.014615396969020367, + 0.01206557173281908, + 0.09390395879745483, + 0.02280806191265583, + 0.07554858177900314, + 0.02714480459690094, + 0.027848871424794197, + 0.027633611112833023, + 0.029325131326913834, + 0.021160388365387917, + 0.04169277101755142, + 0.04501804709434509, + 0.0012201130157336593, + 0.0014200310688465834, + 0.001489941612817347, + 0.0017388378037139773, + 0.004897573497146368, + 0.0016178191872313619, + 0.003571006003767252, + 0.0012131580151617527, + 0.0018576488364487886, + 0.0018377216765657067, + 0.0025013231206685305, + 0.002587159164249897, + 0.0016464818036183715, + 0.0015959276352077723, + 0.004344245884567499, + 0.001900656963698566, + 0.0112331323325634, + 0.011983949691057205, + 0.01498954463750124, + 0.0147560341283679, + 0.01003042608499527, + 0.00914094690233469, + 0.026320399716496468, + 0.0188461784273386, + 0.002811665879562497, + 0.0019356688717380166, + 0.00403430312871933, + 0.0019698976539075375, + 0.003597022034227848, + 0.0031146469991654158, + 0.011166766285896301, + 0.008479210548102856, + 0.008124832063913345, + 0.010973215103149414, + 0.011035550385713577, + 0.008204934187233448, + 0.009912981651723385, + 0.005341235548257828, + 0.009016044437885284, + 0.007074913941323757, + 0.02347029373049736, + 0.017955904826521873, + 0.03399859741330147, + 0.016543108969926834, + 0.016175510361790657, + 0.016662638634443283, + 0.024204712361097336, + 0.014000318013131618, + 0.0018335459753870964, + 0.001713148900307715, + 0.0015144856879487634, + 0.0015585761284455657, + 0.0013407939113676548, + 0.0014827832346782088, + 0.0017868333961814642, + 0.0008910599281080067, + 0.005511398892849684, + 0.0033460301347076893, + 0.007132991217076778, + 0.0031991919968277216, + 0.003289960091933608, + 0.005086155142635107, + 0.0035109722521156073, + 0.0032853628508746624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.14281140267848969 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic shot, initially static, captures two humanoid figures composed entirely of delicate, overlapping pink rose petals, engaging in a graceful courtship dance amidst a surreal, dreamlike landscape. Their intricately layered forms, with petals ranging from soft blush to deep burgundy, create a sense of depth and texture, while their movements are fluid and incredibly graceful. They twirl and embrace, their petal-laden limbs entwined; during their embrace, their forms temporarily merge into a beautiful blossom, individual petals swirling together before gently pulling apart, perfectly reforming their individual forms. The ground beneath their feet is soft, luminous moss, bathing the scene in an ethereal, diffuse light from an unseen source, casting no harsh shadows. As their enchanting ritual continues, the camera begins a slow, seamless crane shot, rising vertically to reveal the vastness of their surroundings. The two figures are part of a vast, blooming garden, where countless other flowers – lilies, tulips and peonies – also take on human shapes and perform a similar, mesmerizing dance, creating a silent, synchronized ballet across a dreamlike terrain, fostering a profound atmosphere of tranquility, magic and romantic wonder.", + "seed": 12531, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_3/index_6281_seed_12531_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2656600184738636, + "vlm_logprob": -0.2656600184738636, + "vlm_yes_no_margin": 1.900000011920929, + "vlm_token_credit": -0.2656600184738636, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03040214255452156 + ], + "opposite_logprob_tokens": [ + -3.905402183532715 + ], + "logprob_reward": -0.03040214255452156, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03040214255452156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12545384466648102, + 0.0483027845621109, + 0.26917195320129395, + 0.07191292196512222, + 0.13361525535583496, + 0.07840763032436371, + 0.07952713966369629, + 0.05897646024823189, + 0.061556439846754074, + 0.07307559996843338 + ], + "credit_map": [ + 0.00906860176473856, + 0.012869315221905708, + 0.010988985188305378, + 0.03816172480583191, + 0.00886490661650896, + 0.007257716730237007, + 0.023616330698132515, + 0.01462625339627266, + 0.004672273062169552, + 0.005596657283604145, + 0.004990801215171814, + 0.010918320156633854, + 0.004643774591386318, + 0.004421974532306194, + 0.00653091398999095, + 0.006528069730848074, + 0.01084652729332447, + 0.02369525469839573, + 0.04215885326266289, + 0.09676094353199005, + 0.01480728667229414, + 0.010616428218781948, + 0.04799577221274376, + 0.02229086123406887, + 0.005702490918338299, + 0.010890588164329529, + 0.007370964623987675, + 0.00888377521187067, + 0.009809769690036774, + 0.008013518527150154, + 0.010734427720308304, + 0.010507392697036266, + 0.006456758826971054, + 0.0058168382383883, + 0.047681208699941635, + 0.009242890402674675, + 0.015648437663912773, + 0.005514694377779961, + 0.037614855915308, + 0.005639567971229553, + 0.006094783544540405, + 0.014958955347537994, + 0.005118582863360643, + 0.010575447231531143, + 0.012365011498332024, + 0.007765660993754864, + 0.013853542506694794, + 0.007675642147660255, + 0.008435647003352642, + 0.00770997442305088, + 0.009079786017537117, + 0.015996765345335007, + 0.008295280858874321, + 0.008401026949286461, + 0.01476460974663496, + 0.006844049785286188, + 0.005226900801062584, + 0.0065615917555987835, + 0.0063094450160861015, + 0.008538045920431614, + 0.006395034026354551, + 0.007083521224558353, + 0.00883246585726738, + 0.01002945750951767, + 0.005844445899128914, + 0.010956761427223682, + 0.009120328351855278, + 0.007245290093123913, + 0.007222888059914112, + 0.00595072889700532, + 0.007996778003871441, + 0.007219221908599138, + 0.008951020427048206, + 0.0076408484019339085, + 0.010379073210060596, + 0.010982213541865349, + 0.007850345224142075, + 0.008344147354364395, + 0.010477556847035885, + 0.008450391702353954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.03040214255452156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280955493450165 + ], + "opposite_logprob_tokens": [ + -2.0178096294403076 + ], + "logprob_reward": -0.14280955493450165, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14280955493450165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09454000741243362, + 0.05677936598658562, + 0.08041618019342422, + 0.03481832519173622, + 0.23336949944496155, + 0.07371379435062408, + 0.12243183702230453, + 0.0645788237452507, + 0.06677126884460449, + 0.17258083820343018 + ], + "credit_map": [ + 0.0102658336982131, + 0.016483403742313385, + 0.011627296917140484, + 0.014937853440642357, + 0.009813218377530575, + 0.008344936184585094, + 0.011456822976469994, + 0.011610642075538635, + 0.004069393500685692, + 0.005812672898173332, + 0.006420008838176727, + 0.010843162424862385, + 0.006488078273832798, + 0.004987792111933231, + 0.009800618514418602, + 0.008357642218470573, + 0.0035369261167943478, + 0.006265866104513407, + 0.01146896742284298, + 0.028620978817343712, + 0.004010647069662809, + 0.0033430641051381826, + 0.0152857955545187, + 0.007883947342634201, + 0.003987234551459551, + 0.0038777892477810383, + 0.004213957116007805, + 0.007044320460408926, + 0.002651222748681903, + 0.003273410489782691, + 0.005302106495946646, + 0.004468286409974098, + 0.014464186504483223, + 0.013879550620913506, + 0.08232678472995758, + 0.015829559415578842, + 0.027919810265302658, + 0.00969622191041708, + 0.058061737567186356, + 0.0111916558817029, + 0.004322613123804331, + 0.010413778014481068, + 0.004500581417232752, + 0.01780976541340351, + 0.009239871986210346, + 0.004725245293229818, + 0.01811661757528782, + 0.004585325717926025, + 0.010471338406205177, + 0.015543082728981972, + 0.013375865295529366, + 0.015587977133691311, + 0.01581697352230549, + 0.016966357827186584, + 0.019688379019498825, + 0.014981876127421856, + 0.0051402803510427475, + 0.0056661986745893955, + 0.008704415522515774, + 0.010056759230792522, + 0.005871312227100134, + 0.010072511620819569, + 0.008850221522152424, + 0.010217133909463882, + 0.00854087807238102, + 0.007752257864922285, + 0.011097047477960587, + 0.006744288839399815, + 0.007779406383633614, + 0.006014135666191578, + 0.008726252242922783, + 0.010117008350789547, + 0.015702322125434875, + 0.019358351826667786, + 0.02835942432284355, + 0.02405218780040741, + 0.016298800706863403, + 0.015574152581393719, + 0.03493545949459076, + 0.018300151452422142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.14280955493450165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865625500679016 + ], + "opposite_logprob_tokens": [ + -3.0486562252044678 + ], + "logprob_reward": -0.04865625500679016, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865625500679016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17312932014465332, + 0.04188985377550125, + 0.2760567367076874, + 0.025322265923023224, + 0.19437871873378754, + 0.06809826195240021, + 0.04689570516347885, + 0.09192312508821487, + 0.02218712493777275, + 0.0601189061999321 + ], + "credit_map": [ + 0.013585072942078114, + 0.013971845619380474, + 0.020716795697808266, + 0.045039914548397064, + 0.012589856050908566, + 0.008816549554467201, + 0.0445614829659462, + 0.013847815804183483, + 0.004537208937108517, + 0.0047782836481928825, + 0.007405353244394064, + 0.007434099912643433, + 0.003803752362728119, + 0.0040150913409888744, + 0.0048075891099870205, + 0.005108474753797054, + 0.011444556526839733, + 0.02062041126191616, + 0.03934526816010475, + 0.10849131643772125, + 0.010739220306277275, + 0.007299124263226986, + 0.050827011466026306, + 0.027289852499961853, + 0.002715661656111479, + 0.00406204117462039, + 0.002771995263174176, + 0.004652062896639109, + 0.002822033828124404, + 0.0022836574353277683, + 0.0033097658306360245, + 0.0027050457429140806, + 0.010391075164079666, + 0.006898283027112484, + 0.0636625811457634, + 0.01579364389181137, + 0.008168503642082214, + 0.006095832213759422, + 0.07364720106124878, + 0.009721592999994755, + 0.0064349425956606865, + 0.01467503048479557, + 0.005521510727703571, + 0.005789481103420258, + 0.013032129034399986, + 0.007349255494773388, + 0.008715750649571419, + 0.006580156274139881, + 0.004778604488819838, + 0.0048577506095170975, + 0.005737203639000654, + 0.0072457860223948956, + 0.006888567935675383, + 0.005110400263220072, + 0.0077688489109277725, + 0.004508544225245714, + 0.006265158299356699, + 0.007338595110923052, + 0.009665096178650856, + 0.012637326493859291, + 0.007623113226145506, + 0.017718501389026642, + 0.012304196134209633, + 0.018371138721704483, + 0.002010506810620427, + 0.0022014863789081573, + 0.00453384080901742, + 0.0033678163308650255, + 0.002510865917429328, + 0.002495374996215105, + 0.0028591882437467575, + 0.0022080461494624615, + 0.0071608866564929485, + 0.011535806581377983, + 0.008317061699926853, + 0.007331993896514177, + 0.007579183205962181, + 0.005347049329429865, + 0.006935231387615204, + 0.005911688320338726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.04865625500679016 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7577775716781616 + ], + "opposite_logprob_tokens": [ + -0.6327775716781616 + ], + "logprob_reward": -0.7577775716781616, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7577775716781616, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04424644634127617, + 0.053098052740097046, + 0.2842286229133606, + 0.02436317875981331, + 0.2015390694141388, + 0.059392787516117096, + 0.04801390320062637, + 0.12643106281757355, + 0.07059735804796219, + 0.08808958530426025 + ], + "credit_map": [ + 0.004058272577822208, + 0.0059074112214148045, + 0.004133354872465134, + 0.009220360778272152, + 0.0048050847835838795, + 0.00336601585149765, + 0.006644933484494686, + 0.006111012306064367, + 0.0038975083734840155, + 0.006527814548462629, + 0.006311068776994944, + 0.007836555130779743, + 0.007167404051870108, + 0.004923971835523844, + 0.008874726481735706, + 0.007559001911431551, + 0.012455880641937256, + 0.027047904208302498, + 0.0358075350522995, + 0.10322164744138718, + 0.015625786036252975, + 0.009899463504552841, + 0.05228168144822121, + 0.027888724580407143, + 0.002569820499047637, + 0.0035427489783614874, + 0.0029202031437307596, + 0.003746010595932603, + 0.0023005432449281216, + 0.0025389099027961493, + 0.004073204006999731, + 0.002671740483492613, + 0.010009247809648514, + 0.013115267269313335, + 0.0652168020606041, + 0.016421126201748848, + 0.025937948375940323, + 0.010127000510692596, + 0.051239922642707825, + 0.009471750818192959, + 0.004213918000459671, + 0.006650957744568586, + 0.005165833979845047, + 0.013059196062386036, + 0.005139004439115524, + 0.007130663841962814, + 0.009255933575332165, + 0.008777273818850517, + 0.00582753773778677, + 0.006044316105544567, + 0.006307624280452728, + 0.005981722380965948, + 0.005566813051700592, + 0.005797339603304863, + 0.007472844794392586, + 0.0050157043151557446, + 0.012580826878547668, + 0.019317131489515305, + 0.014238917268812656, + 0.01697334088385105, + 0.01874428801238537, + 0.012483248487114906, + 0.017994720488786697, + 0.014098594896495342, + 0.007299208547919989, + 0.014492892660200596, + 0.009134541265666485, + 0.008867732249200344, + 0.008719406090676785, + 0.004882078152149916, + 0.009646215476095676, + 0.007555282209068537, + 0.010803611017763615, + 0.00821846816688776, + 0.009381896816194057, + 0.014538273215293884, + 0.010270413011312485, + 0.010704006999731064, + 0.01333944033831358, + 0.010833470150828362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.7577775716781616 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.348654568195343 + ], + "opposite_logprob_tokens": [ + -1.2236545085906982 + ], + "logprob_reward": -0.348654568195343, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.348654568195343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04706942290067673, + 0.016726821660995483, + 0.09331947565078735, + 0.02784854918718338, + 0.49689391255378723, + 0.06406524777412415, + 0.13928882777690887, + 0.0647149309515953, + 0.02874869853258133, + 0.021324053406715393 + ], + "credit_map": [ + 0.0032587561290711164, + 0.00499627785757184, + 0.0035665989853441715, + 0.013975742273032665, + 0.003584513906389475, + 0.0031867632642388344, + 0.009862799197435379, + 0.004637969192117453, + 0.0022469055838882923, + 0.002906089648604393, + 0.001397404121235013, + 0.002183595672249794, + 0.0016901218332350254, + 0.0015975241549313068, + 0.002358239609748125, + 0.0023469424340873957, + 0.0035027186386287212, + 0.007950124330818653, + 0.01418967917561531, + 0.03495682403445244, + 0.004408461041748524, + 0.003147316398099065, + 0.016477396711707115, + 0.008686961606144905, + 0.0026225121691823006, + 0.00391674367710948, + 0.0035023328382521868, + 0.003207687521353364, + 0.0031075789593160152, + 0.002744549885392189, + 0.005369795020669699, + 0.003377350978553295, + 0.02670661360025406, + 0.020895855501294136, + 0.18200895190238953, + 0.03308754041790962, + 0.054351259022951126, + 0.019421927630901337, + 0.13600096106529236, + 0.024420851841568947, + 0.002413627225905657, + 0.006662912201136351, + 0.004229717422276735, + 0.02179783023893833, + 0.005845080595463514, + 0.0033467186149209738, + 0.016341974958777428, + 0.0034273904748260975, + 0.011612591333687305, + 0.016340753063559532, + 0.009499982930719852, + 0.014768376015126705, + 0.016035081818699837, + 0.030765023082494736, + 0.019087374210357666, + 0.021179642528295517, + 0.004807412624359131, + 0.003824136219918728, + 0.005478285253047943, + 0.01002112403512001, + 0.005024043843150139, + 0.012094938196241856, + 0.008771108463406563, + 0.01469388883560896, + 0.003265125211328268, + 0.00518769072368741, + 0.0036593477707356215, + 0.003444408066570759, + 0.0036694190930575132, + 0.0020896282512694597, + 0.003751002484932542, + 0.003682077396661043, + 0.003060530871152878, + 0.0019839443266391754, + 0.0033182757906615734, + 0.003128268290311098, + 0.002023532288148999, + 0.002298078266903758, + 0.0022957781329751015, + 0.0032156461384147406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.348654568195343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "cinematic shot, initially static, captures two humanoid figures composed entirely of delicate, overlapping pink rose petals, engaging in a graceful courtship dance amidst a surreal, dreamlike landscape. Their intricately layered forms, with petals ranging from soft blush to deep burgundy, create a sense of depth and texture, while their movements are fluid and incredibly graceful. They twirl and embrace, their petal-laden limbs entwined; during their embrace, their forms temporarily merge into a beautiful blossom, individual petals swirling together before gently pulling apart, perfectly reforming their individual forms. The ground beneath their feet is soft, luminous moss, bathing the scene in an ethereal, diffuse light from an unseen source, casting no harsh shadows. As their enchanting ritual continues, the camera begins a slow, seamless crane shot, rising vertically to reveal the vastness of their surroundings. The two figures are part of a vast, blooming garden, where countless other flowers – lilies, tulips and peonies – also take on human shapes and perform a similar, mesmerizing dance, creating a silent, synchronized ballet across a dreamlike terrain, fostering a profound atmosphere of tranquility, magic and romantic wonder.", + "seed": 12532, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_3/index_6281_seed_12532_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1193536464124918, + "vlm_logprob": -0.1193536464124918, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.1193536464124918, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02511746622622013 + ], + "opposite_logprob_tokens": [ + -4.40011739730835 + ], + "logprob_reward": -0.02511746622622013, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02511746622622013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09549140185117722, + 0.08663170784711838, + 0.0292800422757864, + 0.07185040414333344, + 0.11329708993434906, + 0.08270084857940674, + 0.08848613500595093, + 0.04943717271089554, + 0.3279295265674591, + 0.05489565059542656 + ], + "credit_map": [ + 0.008403677493333817, + 0.01601681299507618, + 0.009752850048244, + 0.012323160655796528, + 0.015425409190356731, + 0.009632259607315063, + 0.012895237654447556, + 0.011042002588510513, + 0.007332934532314539, + 0.011736358515918255, + 0.012792503461241722, + 0.012850255705416203, + 0.01177526917308569, + 0.005313625559210777, + 0.015577388927340508, + 0.00925337802618742, + 0.0019717933610081673, + 0.003675607033073902, + 0.0032428873237222433, + 0.005069599952548742, + 0.004963060840964317, + 0.00230072601698339, + 0.005052625667303801, + 0.0030037444084882736, + 0.007813971489667892, + 0.013116305693984032, + 0.01583166792988777, + 0.0057587879709899426, + 0.005294333212077618, + 0.006778253708034754, + 0.008411824703216553, + 0.008845263160765171, + 0.008697330951690674, + 0.018482252955436707, + 0.017485784366726875, + 0.020720230415463448, + 0.012716121971607208, + 0.012275989167392254, + 0.013705417513847351, + 0.009213965386152267, + 0.006533282808959484, + 0.014766174368560314, + 0.010762931779026985, + 0.012617522850632668, + 0.010012230835855007, + 0.007940772920846939, + 0.012069456279277802, + 0.007998484186828136, + 0.0052827633917331696, + 0.006742626894265413, + 0.004384421743452549, + 0.029753658920526505, + 0.004977541510015726, + 0.006594299338757992, + 0.024516887962818146, + 0.006233936175704002, + 0.006769654341042042, + 0.005369926802814007, + 0.0054583377204835415, + 0.005797151010483503, + 0.006085466127842665, + 0.009135446511209011, + 0.006164712831377983, + 0.004656480625271797, + 0.042416803538799286, + 0.048883069306612015, + 0.026821589097380638, + 0.04875498265028, + 0.04690660536289215, + 0.044326722621917725, + 0.04326377809047699, + 0.026555968448519707, + 0.006588035263121128, + 0.006492006126791239, + 0.0077018048614263535, + 0.006515230983495712, + 0.008890163153409958, + 0.00814457330852747, + 0.005720591638237238, + 0.004843253176659346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.02511746622622013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904742449522018 + ], + "opposite_logprob_tokens": [ + -2.464047431945801 + ], + "logprob_reward": -0.08904742449522018, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904742449522018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1387217938899994, + 0.08386389166116714, + 0.05176444351673126, + 0.10514863580465317, + 0.10735585540533066, + 0.09334060549736023, + 0.18165786564350128, + 0.1478530615568161, + 0.02867146022617817, + 0.061622388660907745 + ], + "credit_map": [ + 0.011132743209600449, + 0.02104908786714077, + 0.018496103584766388, + 0.01407778263092041, + 0.019489189609885216, + 0.011587796732783318, + 0.028108153492212296, + 0.014780943281948566, + 0.007158046122640371, + 0.012505365535616875, + 0.012334524653851986, + 0.013279331848025322, + 0.009337598457932472, + 0.006147691980004311, + 0.013491267338395119, + 0.009610069915652275, + 0.0040136040188372135, + 0.006292169447988272, + 0.007020092103630304, + 0.007281783036887646, + 0.0071306354366242886, + 0.005032026674598455, + 0.007868492975831032, + 0.007125642616301775, + 0.009406138211488724, + 0.019640928134322166, + 0.029226455837488174, + 0.010442749597132206, + 0.007936178706586361, + 0.009171479381620884, + 0.011146029457449913, + 0.008178677409887314, + 0.007810165174305439, + 0.01688111387193203, + 0.01536625623703003, + 0.0219077430665493, + 0.014536149799823761, + 0.010443966835737228, + 0.01190018281340599, + 0.008510280400514603, + 0.007319557014852762, + 0.01936783827841282, + 0.010624771006405354, + 0.012310161255300045, + 0.014641874469816685, + 0.008377589285373688, + 0.012079252861440182, + 0.008619560860097408, + 0.010025771334767342, + 0.015676986426115036, + 0.00740852952003479, + 0.06633816659450531, + 0.00975807849317789, + 0.007633520755916834, + 0.054207880049943924, + 0.01060892641544342, + 0.016566244885325432, + 0.012497277930378914, + 0.01879817061126232, + 0.02027387171983719, + 0.01284834649413824, + 0.0171134565025568, + 0.02220352366566658, + 0.027552170678973198, + 0.0035613675136119127, + 0.00513205723837018, + 0.002203922951593995, + 0.004814524669200182, + 0.003214389318600297, + 0.004624794237315655, + 0.0034654256887733936, + 0.0016549801221117377, + 0.007797492668032646, + 0.0056229219771921635, + 0.008977373130619526, + 0.006210708059370518, + 0.006653922609984875, + 0.007443474605679512, + 0.010242805816233158, + 0.00867368932813406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.08904742449522018 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369775414466858 + ], + "opposite_logprob_tokens": [ + -3.4086978435516357 + ], + "logprob_reward": -0.03369775414466858, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369775414466858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18574468791484833, + 0.09100734442472458, + 0.05435815081000328, + 0.07354426383972168, + 0.09662100672721863, + 0.19952522218227386, + 0.07012391090393066, + 0.05315204709768295, + 0.07568272203207016, + 0.10024061053991318 + ], + "credit_map": [ + 0.015375985763967037, + 0.030969789251685143, + 0.01652463898062706, + 0.021101800724864006, + 0.03343718871474266, + 0.020869800820946693, + 0.028362857177853584, + 0.019102632999420166, + 0.005488147959113121, + 0.013577266596257687, + 0.010900316759943962, + 0.013497628271579742, + 0.011789237149059772, + 0.00758916512131691, + 0.016087286174297333, + 0.012078301049768925, + 0.006478210911154747, + 0.008007069118320942, + 0.0090226661413908, + 0.006450153887271881, + 0.006133579649031162, + 0.004165319260209799, + 0.007165222428739071, + 0.0069359284825623035, + 0.00825138483196497, + 0.012176180258393288, + 0.009456337429583073, + 0.00752053689211607, + 0.006747140083462, + 0.01360473781824112, + 0.008267536759376526, + 0.007520417217165232, + 0.006370453163981438, + 0.014412264339625835, + 0.013954497873783112, + 0.01657324843108654, + 0.012387753464281559, + 0.009478983469307423, + 0.014717241749167442, + 0.008726566098630428, + 0.015498535707592964, + 0.035681311041116714, + 0.028297526761889458, + 0.023532595485448837, + 0.031173648312687874, + 0.019121307879686356, + 0.022457454353570938, + 0.023762863129377365, + 0.004132171161472797, + 0.005398490931838751, + 0.00317372870631516, + 0.02418898604810238, + 0.003918871283531189, + 0.0035547586157917976, + 0.02212698385119438, + 0.0036299226339906454, + 0.0060760523192584515, + 0.005034668371081352, + 0.005434627179056406, + 0.006943844724446535, + 0.006016543600708246, + 0.008296363987028599, + 0.00891527347266674, + 0.006434671580791473, + 0.009436315856873989, + 0.012447002343833447, + 0.006223318632692099, + 0.01037645060569048, + 0.010220441035926342, + 0.010436507873237133, + 0.00914300512522459, + 0.007399688474833965, + 0.012772670947015285, + 0.009495017118752003, + 0.017509346827864647, + 0.009537927806377411, + 0.013262199237942696, + 0.01139174122363329, + 0.014166540466248989, + 0.01210517156869173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.03369775414466858 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486020565032959 + ], + "opposite_logprob_tokens": [ + -1.223602056503296 + ], + "logprob_reward": -0.3486020565032959, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486020565032959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10267006605863571, + 0.07210459560155869, + 0.07975073158740997, + 0.12639082968235016, + 0.09950816631317139, + 0.13366997241973877, + 0.10326656699180603, + 0.0616125725209713, + 0.19056624174118042, + 0.030460229143500328 + ], + "credit_map": [ + 0.008267657831311226, + 0.015927094966173172, + 0.012198623269796371, + 0.011039751581847668, + 0.014923696406185627, + 0.007208151277154684, + 0.020486650988459587, + 0.012618440203368664, + 0.005800873506814241, + 0.012151595205068588, + 0.007820109836757183, + 0.010699166916310787, + 0.009820083156228065, + 0.0075147030875086784, + 0.010240173898637295, + 0.008057896979153156, + 0.007903272286057472, + 0.01075681671500206, + 0.00985565036535263, + 0.009605237282812595, + 0.012400807812809944, + 0.010044670663774014, + 0.011430416256189346, + 0.007753861602395773, + 0.013866458088159561, + 0.023178698495030403, + 0.021991178393363953, + 0.010899300687015057, + 0.009886395186185837, + 0.01773373782634735, + 0.014135671779513359, + 0.014699386432766914, + 0.007196206599473953, + 0.01534198597073555, + 0.015773074701428413, + 0.019111167639493942, + 0.012636610306799412, + 0.008923212997615337, + 0.011179150082170963, + 0.009346753358840942, + 0.012979828752577305, + 0.01962650939822197, + 0.01736096478998661, + 0.019108230248093605, + 0.015515359118580818, + 0.019600654020905495, + 0.015385296195745468, + 0.01409312803298235, + 0.007256947923451662, + 0.015118066221475601, + 0.005423626396805048, + 0.028917232528328896, + 0.008945919573307037, + 0.006320137996226549, + 0.022972585633397102, + 0.008312047459185123, + 0.006355370860546827, + 0.005666240584105253, + 0.0054770540446043015, + 0.0075996033847332, + 0.010032451711595058, + 0.0075173950754106045, + 0.009992026723921299, + 0.008972431533038616, + 0.02745380438864231, + 0.028856946155428886, + 0.016011394560337067, + 0.028309796005487442, + 0.023663267493247986, + 0.028829557821154594, + 0.02317110449075699, + 0.014270367100834846, + 0.004596461541950703, + 0.003950515761971474, + 0.003763059386983514, + 0.0035831641871482134, + 0.0032568960450589657, + 0.004779288545250893, + 0.0034180725924670696, + 0.003112772246822715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.3486020565032959 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003035306930542 + ], + "opposite_logprob_tokens": [ + -2.3503036499023438 + ], + "logprob_reward": -0.1003035306930542, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003035306930542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07648802548646927, + 0.075468510389328, + 0.034656498581171036, + 0.02809574268758297, + 0.09596043825149536, + 0.11859919875860214, + 0.38130274415016174, + 0.043648611754179, + 0.12959930300712585, + 0.016180885955691338 + ], + "credit_map": [ + 0.0057263742201030254, + 0.010618604719638824, + 0.006145824678242207, + 0.008227926678955555, + 0.014019844122231007, + 0.009354730136692524, + 0.013848291710019112, + 0.008546430617570877, + 0.005791287869215012, + 0.011605624109506607, + 0.011538958176970482, + 0.011952802538871765, + 0.00858452171087265, + 0.0051034861244261265, + 0.01369145605713129, + 0.007200376596301794, + 0.003338998882099986, + 0.005691568832844496, + 0.004154632333666086, + 0.0042868745513260365, + 0.002672036411240697, + 0.005040231626480818, + 0.0030361663084477186, + 0.00643598847091198, + 0.002889226656407118, + 0.004317013081163168, + 0.004155306611210108, + 0.0034400939475744963, + 0.003309515304863453, + 0.0037961616180837154, + 0.0033394761849194765, + 0.0028489476535469294, + 0.008142531849443913, + 0.015588674694299698, + 0.016481462866067886, + 0.01700047217309475, + 0.009747184813022614, + 0.009769557975232601, + 0.010866827331483364, + 0.008363734930753708, + 0.008838104084134102, + 0.025292305275797844, + 0.011673405766487122, + 0.01740218885242939, + 0.0185532346367836, + 0.011707750149071217, + 0.015351049602031708, + 0.009781165048480034, + 0.020941007882356644, + 0.0323096364736557, + 0.015607796609401703, + 0.13736805319786072, + 0.019864078611135483, + 0.015590261667966843, + 0.11992278695106506, + 0.01969909854233265, + 0.005442868918180466, + 0.003167776856571436, + 0.006208366714417934, + 0.006213529966771603, + 0.003338563023135066, + 0.005238841287791729, + 0.004829995334148407, + 0.009208671748638153, + 0.015547802671790123, + 0.0205873791128397, + 0.00965789146721363, + 0.01970260590314865, + 0.02018868736922741, + 0.017826268449425697, + 0.01662530191242695, + 0.009463372640311718, + 0.0022407425567507744, + 0.001975136809051037, + 0.0024257467593997717, + 0.0018324046395719051, + 0.0030693586450070143, + 0.0016683923313394189, + 0.0016717907274141908, + 0.0012973124394193292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.1003035306930542 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic shot, initially static, captures two humanoid figures composed entirely of delicate, overlapping pink rose petals, engaging in a graceful courtship dance amidst a surreal, dreamlike landscape. Their intricately layered forms, with petals ranging from soft blush to deep burgundy, create a sense of depth and texture, while their movements are fluid and incredibly graceful. They twirl and embrace, their petal-laden limbs entwined; during their embrace, their forms temporarily merge into a beautiful blossom, individual petals swirling together before gently pulling apart, perfectly reforming their individual forms. The ground beneath their feet is soft, luminous moss, bathing the scene in an ethereal, diffuse light from an unseen source, casting no harsh shadows. As their enchanting ritual continues, the camera begins a slow, seamless crane shot, rising vertically to reveal the vastness of their surroundings. The two figures are part of a vast, blooming garden, where countless other flowers – lilies, tulips and peonies – also take on human shapes and perform a similar, mesmerizing dance, creating a silent, synchronized ballet across a dreamlike terrain, fostering a profound atmosphere of tranquility, magic and romantic wonder.", + "seed": 12533, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_3/index_6281_seed_12533_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15654038675129414, + "vlm_logprob": -0.15654038675129414, + "vlm_yes_no_margin": 2.449999970197678, + "vlm_token_credit": -0.15654038675129414, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02288106083869934 + ], + "opposite_logprob_tokens": [ + -4.272881031036377 + ], + "logprob_reward": -0.02288106083869934, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02288106083869934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10346516221761703, + 0.04723837599158287, + 0.041074808686971664, + 0.0646485686302185, + 0.28722137212753296, + 0.1528891772031784, + 0.08283926546573639, + 0.156845822930336, + 0.030127836391329765, + 0.033649660646915436 + ], + "credit_map": [ + 0.01013942901045084, + 0.016657806932926178, + 0.009685954079031944, + 0.012147966772317886, + 0.01386775728315115, + 0.013726101256906986, + 0.017096804454922676, + 0.010143331252038479, + 0.005458187777549028, + 0.00414600083604455, + 0.008984050713479519, + 0.004114020150154829, + 0.005404007155448198, + 0.005875945556908846, + 0.005649042781442404, + 0.007607113569974899, + 0.003927014768123627, + 0.004242675844579935, + 0.005547911860048771, + 0.004983976949006319, + 0.006041660904884338, + 0.004434624686837196, + 0.0037641774397343397, + 0.008132762275636196, + 0.00577889196574688, + 0.009251078590750694, + 0.005512799136340618, + 0.01280688401311636, + 0.008804535493254662, + 0.00897788442671299, + 0.007099854294210672, + 0.006416634190827608, + 0.015133889392018318, + 0.11644675582647324, + 0.08830380439758301, + 0.01534140482544899, + 0.023499595001339912, + 0.006952913943678141, + 0.013015580363571644, + 0.00852741114795208, + 0.014002648182213306, + 0.011013285256922245, + 0.03891311585903168, + 0.016330959275364876, + 0.03899815306067467, + 0.01040672417730093, + 0.015653682872653008, + 0.007570603862404823, + 0.011495775543153286, + 0.011816397309303284, + 0.010584981180727482, + 0.010097445920109749, + 0.009419204667210579, + 0.011970150284469128, + 0.009316345676779747, + 0.008138950914144516, + 0.0229807011783123, + 0.04591827839612961, + 0.018111474812030792, + 0.014915553852915764, + 0.013144198805093765, + 0.014543533325195312, + 0.018255839124321938, + 0.008976232260465622, + 0.003469279268756509, + 0.003778203157708049, + 0.004934840835630894, + 0.0028961333446204662, + 0.0039812359027564526, + 0.003517360892146826, + 0.004134390503168106, + 0.0034163908567279577, + 0.003976389765739441, + 0.0036945627070963383, + 0.003824080340564251, + 0.005063085351139307, + 0.005349823739379644, + 0.0036277405451983213, + 0.004305078648030758, + 0.0038088939618319273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 5 + ], + "selected_reward": -0.02288106083869934 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294480413198471 + ], + "opposite_logprob_tokens": [ + -2.2379448413848877 + ], + "logprob_reward": -0.11294480413198471, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294480413198471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04809265211224556, + 0.05369481444358826, + 0.0756533145904541, + 0.03309508413076401, + 0.3641628921031952, + 0.0453685000538826, + 0.05936536192893982, + 0.20010487735271454, + 0.04153807833790779, + 0.07892441749572754 + ], + "credit_map": [ + 0.005600763019174337, + 0.006992526818066835, + 0.007148537319153547, + 0.005892119836062193, + 0.005168440286070108, + 0.005140815861523151, + 0.006391691043972969, + 0.005757762119174004, + 0.005748439580202103, + 0.008008021861314774, + 0.005598427727818489, + 0.00517028383910656, + 0.007092414423823357, + 0.006822860799729824, + 0.0054512033239007, + 0.009803172200918198, + 0.008636762388050556, + 0.011008170433342457, + 0.005901514086872339, + 0.011219208128750324, + 0.009102639742195606, + 0.010594869963824749, + 0.01173948124051094, + 0.007450679782778025, + 0.003186616348102689, + 0.005293326918035746, + 0.0026755286380648613, + 0.005762708373367786, + 0.004724207799881697, + 0.003624788485467434, + 0.003897675545886159, + 0.003930235747247934, + 0.019104128703475, + 0.14530432224273682, + 0.11259368807077408, + 0.020067578181624413, + 0.029033828526735306, + 0.010551370680332184, + 0.015962090343236923, + 0.011545940302312374, + 0.004418053198605776, + 0.003655733773484826, + 0.010979299433529377, + 0.004903342574834824, + 0.011143217794597149, + 0.0025747220497578382, + 0.0053795683197677135, + 0.0023145689629018307, + 0.00756805157288909, + 0.007113106548786163, + 0.007525596302002668, + 0.00788557343184948, + 0.007836833596229553, + 0.007826962508261204, + 0.007027254905551672, + 0.006581994239240885, + 0.03285757824778557, + 0.06382329016923904, + 0.019412387162446976, + 0.020181164145469666, + 0.01715737394988537, + 0.012479646131396294, + 0.02301567792892456, + 0.011177781969308853, + 0.005593420006334782, + 0.007024995516985655, + 0.003679667366668582, + 0.004687167704105377, + 0.006575310602784157, + 0.003541115904226899, + 0.005862516816705465, + 0.004573888145387173, + 0.008992709219455719, + 0.009799173101782799, + 0.008167464286088943, + 0.012185420840978622, + 0.013736456632614136, + 0.009994476102292538, + 0.009045750834047794, + 0.007002974394708872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "selected_reward": -0.11294480413198471 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029800983145833015 + ], + "opposite_logprob_tokens": [ + -3.5298008918762207 + ], + "logprob_reward": -0.029800983145833015, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029800983145833015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10401271283626556, + 0.05852315574884415, + 0.08600519597530365, + 0.2218371480703354, + 0.05364851653575897, + 0.050762347877025604, + 0.11025149375200272, + 0.20581981539726257, + 0.08173996210098267, + 0.02739970199763775 + ], + "credit_map": [ + 0.014750104397535324, + 0.01423682551831007, + 0.014075050130486488, + 0.010752877220511436, + 0.011034630239009857, + 0.01517841499298811, + 0.00955576915293932, + 0.014429028145968914, + 0.006284629460424185, + 0.008367790840566158, + 0.007681783754378557, + 0.007229945622384548, + 0.0073408205062150955, + 0.006848678924143314, + 0.006758576259016991, + 0.008010923862457275, + 0.010685691609978676, + 0.012971064075827599, + 0.0071310196071863174, + 0.012415061704814434, + 0.010300429537892342, + 0.011439753696322441, + 0.012289615347981453, + 0.008772546425461769, + 0.03151837736368179, + 0.020865490660071373, + 0.02765120193362236, + 0.040697626769542694, + 0.03199775516986847, + 0.023648938164114952, + 0.023527242243289948, + 0.021930504590272903, + 0.0029067303985357285, + 0.01719892956316471, + 0.01357357855886221, + 0.006027287803590298, + 0.004613696597516537, + 0.0024936117697507143, + 0.003375865053385496, + 0.0034588139969855547, + 0.006073359400033951, + 0.006116594187915325, + 0.004864040296524763, + 0.005525721237063408, + 0.011758890002965927, + 0.006821414455771446, + 0.00603953143581748, + 0.0035627938341349363, + 0.018093319609761238, + 0.011036766692996025, + 0.02267749421298504, + 0.014595572836697102, + 0.01027185469865799, + 0.012423497624695301, + 0.011487746611237526, + 0.009665228426456451, + 0.030953610315918922, + 0.067570760846138, + 0.02001858502626419, + 0.019963664934039116, + 0.015939876437187195, + 0.017458762973546982, + 0.023358119651675224, + 0.010556423105299473, + 0.01350812055170536, + 0.010809310711920261, + 0.007420761976391077, + 0.010068503208458424, + 0.01039072498679161, + 0.006334513425827026, + 0.018163250759243965, + 0.0050447722896933556, + 0.0036947433836758137, + 0.0038082674145698547, + 0.0026514832861721516, + 0.002714429749175906, + 0.00535558070987463, + 0.00324213202111423, + 0.003139989683404565, + 0.0027930729556828737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.029800983145833015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742565453052521 + ], + "opposite_logprob_tokens": [ + -0.9742565155029297 + ], + "logprob_reward": -0.4742565453052521, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.4742565453052521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03614659234881401, + 0.046246565878391266, + 0.17229610681533813, + 0.04920715466141701, + 0.15888206660747528, + 0.13198848068714142, + 0.10556971281766891, + 0.2869178354740143, + 0.007108947262167931, + 0.005636592395603657 + ], + "credit_map": [ + 0.004551074467599392, + 0.005373239051550627, + 0.0037670221645385027, + 0.004274056293070316, + 0.0038668678607791662, + 0.0038069526199251413, + 0.005591931287199259, + 0.004915447905659676, + 0.004514256026595831, + 0.007169931195676327, + 0.00576357264071703, + 0.005798223428428173, + 0.00597989559173584, + 0.006010228302329779, + 0.0052839298732578754, + 0.005726528353989124, + 0.021487096324563026, + 0.020256521180272102, + 0.016036834567785263, + 0.030687551945447922, + 0.01725752092897892, + 0.028617795556783676, + 0.020474636927247047, + 0.017478128895163536, + 0.004922926425933838, + 0.007838030345737934, + 0.0037027716170996428, + 0.008908741176128387, + 0.007080982439219952, + 0.005970069672912359, + 0.005703455302864313, + 0.005080176517367363, + 0.008268608711659908, + 0.06229247897863388, + 0.04739120230078697, + 0.010652094148099422, + 0.013042590580880642, + 0.004850683733820915, + 0.0072507006116211414, + 0.005133711267262697, + 0.0128129618242383, + 0.011244424618780613, + 0.0330519936978817, + 0.01556330919265747, + 0.029794903472065926, + 0.008752193301916122, + 0.013934715650975704, + 0.006833973340690136, + 0.015264157205820084, + 0.014223323203623295, + 0.014691783115267754, + 0.011983315460383892, + 0.01178811863064766, + 0.015328151173889637, + 0.01165872160345316, + 0.010632146149873734, + 0.04615936055779457, + 0.09169843047857285, + 0.029561081901192665, + 0.02737506851553917, + 0.022335126996040344, + 0.02153523452579975, + 0.03321019560098648, + 0.015043298713862896, + 0.0014775219606235623, + 0.00045468812459148467, + 0.0005289107211865485, + 0.0009277256322093308, + 0.0006486372440122068, + 0.0005814984906464815, + 0.0020736821461468935, + 0.0004162830300629139, + 0.0007103986572474241, + 0.0006148413522168994, + 0.0008216178393922746, + 0.0006434823735617101, + 0.0008660607854835689, + 0.0006372033967636526, + 0.0006403059232980013, + 0.0007026820094324648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.4742565453052521 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the two main figures composed entirely of overlapping pink rose petals ranging from soft blush to deep burgundy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the figures twirl and embrace, with their forms temporarily merging into a single blossom before pulling apart?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the ground beneath the figures soft, luminous moss casting no harsh shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin as static and then execute a slow, seamless vertical crane shot rising upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there other humanoid figures shaped like lilies, tulips, and peonies performing a synchronized dance in the background?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281854033470154 + ], + "opposite_logprob_tokens": [ + -2.0178184509277344 + ], + "logprob_reward": -0.14281854033470154, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14281854033470154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04791092500090599, + 0.01886778511106968, + 0.08834109455347061, + 0.03861360251903534, + 0.3335481882095337, + 0.03754911571741104, + 0.12837952375411987, + 0.12018124014139175, + 0.1561332494020462, + 0.030475229024887085 + ], + "credit_map": [ + 0.005941626150161028, + 0.008000691421329975, + 0.006875674705952406, + 0.005281535908579826, + 0.004658956546336412, + 0.005009577609598637, + 0.007372940890491009, + 0.0047699278220534325, + 0.0031296624802052975, + 0.0022929732222110033, + 0.002214473905041814, + 0.0019303977023810148, + 0.0017484872369095683, + 0.0021839849650859833, + 0.002315864432603121, + 0.003051942680031061, + 0.009024109691381454, + 0.007772755343466997, + 0.008234226144850254, + 0.01696614921092987, + 0.008946362882852554, + 0.015562728978693485, + 0.009668054059147835, + 0.012166713364422321, + 0.005309073254466057, + 0.0051982165314257145, + 0.0032994155772030354, + 0.0066131120547652245, + 0.00517078023403883, + 0.005108742509037256, + 0.004681360442191362, + 0.0032329028472304344, + 0.01726890169084072, + 0.13130436837673187, + 0.10194481164216995, + 0.018536917865276337, + 0.026599425822496414, + 0.01069071888923645, + 0.01593347080051899, + 0.011269563809037209, + 0.00326891359873116, + 0.002759010763838887, + 0.006677941419184208, + 0.004187409766018391, + 0.009551310911774635, + 0.0046064592897892, + 0.0038665279280394316, + 0.0026315459981560707, + 0.019473448395729065, + 0.018230577930808067, + 0.0213360246270895, + 0.013510792516171932, + 0.01202469877898693, + 0.018178172409534454, + 0.012467185035347939, + 0.013158641755580902, + 0.015315231867134571, + 0.02916419319808483, + 0.016211139038205147, + 0.012854580767452717, + 0.01329781860113144, + 0.01047274935990572, + 0.012762786820530891, + 0.010102744214236736, + 0.02796461619436741, + 0.01747334934771061, + 0.012520825490355492, + 0.014533547684550285, + 0.015948230400681496, + 0.013395513407886028, + 0.033618588000535965, + 0.020678572356700897, + 0.0036909114569425583, + 0.0029079953674227, + 0.005408726632595062, + 0.002567045856267214, + 0.004574317950755358, + 0.004610927775502205, + 0.0026650531217455864, + 0.00405025202780962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.14281854033470154 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_4/index_tensor([6499, 6499])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_4/index_tensor([6499, 6499])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fc0483078469e98738ce255e99649223345b6589 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_4/index_tensor([6499, 6499])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angle close-up frames a crystal-clear cylindrical glass filled with perfectly still, transparent water, set against a dark, minimalist backdrop. At the bottom of the glass, a solid flake of white soapstone begins to dissolve violently. Instantly, a cloud of giant silver bubbles erupts from the entire surface, forming a shimmering vertical stream that rapidly rises through the water. As this continuous stream of bubbles reaches the surface, they accumulate and coalesce, meticulously forming an expanding layer of dense, pure white foam. The camera captures the dynamic contrast between the chaotic upward surge of bubbles and the steady, deliberate growth of a creamy layer of foam, which gradually thickens and covers the entire surface of the water. All illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the pristine texture of the foam, creating a focused and captivating atmosphere of vibrant transformation.", + "seed": 28418, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_4/index_6499_seed_28418_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.035984650859609245, + "vlm_logprob": -0.035984650859609245, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.035984650859609245, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02799382619559765 + ], + "opposite_logprob_tokens": [ + -4.652993679046631 + ], + "logprob_reward": -0.02799382619559765, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.02799382619559765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21260535717010498, + 0.18201890587806702, + 0.047536034137010574, + 0.08911501616239548, + 0.0793331116437912, + 0.0524611622095108, + 0.057481274008750916, + 0.04832814633846283, + 0.10436628758907318, + 0.12675471603870392 + ], + "credit_map": [ + 0.022899670526385307, + 0.03822765126824379, + 0.03683207556605339, + 0.018967511132359505, + 0.04913187772035599, + 0.01280506607145071, + 0.020813465118408203, + 0.012928038835525513, + 0.02849338948726654, + 0.027103928849101067, + 0.027828998863697052, + 0.02907559834420681, + 0.02528391219675541, + 0.00873412936925888, + 0.023151565343141556, + 0.012347393669188023, + 0.006929223425686359, + 0.007997297681868076, + 0.006013515871018171, + 0.0040535274893045425, + 0.005713579710572958, + 0.005501049570739269, + 0.004828541539609432, + 0.0064993007108569145, + 0.01593802496790886, + 0.011354425922036171, + 0.013038653880357742, + 0.011053439229726791, + 0.01047499105334282, + 0.009775289334356785, + 0.008989540860056877, + 0.008490651845932007, + 0.012178641743957996, + 0.010698443278670311, + 0.012157878838479519, + 0.012000647373497486, + 0.008300983346998692, + 0.007650833576917648, + 0.009956012479960918, + 0.006389673333615065, + 0.00719616562128067, + 0.005119627341628075, + 0.007512920070439577, + 0.009368257597088814, + 0.004143426660448313, + 0.005647291429340839, + 0.007875221781432629, + 0.0055982572957873344, + 0.006042456720024347, + 0.00547470897436142, + 0.008119489066302776, + 0.008647465147078037, + 0.006425295956432819, + 0.006894578225910664, + 0.008841977454721928, + 0.007035307586193085, + 0.005786390043795109, + 0.00637428043410182, + 0.005918774753808975, + 0.005949618294835091, + 0.0055196285247802734, + 0.004285983741283417, + 0.007412412203848362, + 0.007081059738993645, + 0.007430979050695896, + 0.030329691246151924, + 0.011855288408696651, + 0.007197954226285219, + 0.018384834751486778, + 0.007263643201440573, + 0.013662775047123432, + 0.00824112631380558, + 0.015042845159769058, + 0.021119175478816032, + 0.006527964491397142, + 0.01058986783027649, + 0.0338263176381588, + 0.013936474919319153, + 0.013634410686790943, + 0.01207766868174076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.02799382619559765 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790054127573967 + ], + "opposite_logprob_tokens": [ + -2.579005479812622 + ], + "logprob_reward": -0.0790054127573967, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790054127573967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2906103730201721, + 0.13717827200889587, + 0.08005126565694809, + 0.07696567475795746, + 0.12024425715208054, + 0.046878714114427567, + 0.04124007746577263, + 0.06384649872779846, + 0.03164023160934448, + 0.11134469509124756 + ], + "credit_map": [ + 0.04435213655233383, + 0.049219127744436264, + 0.04395274817943573, + 0.02667204476892948, + 0.05763033404946327, + 0.02248590998351574, + 0.02301442064344883, + 0.023283636197447777, + 0.017207643017172813, + 0.01978841982781887, + 0.01807185262441635, + 0.022791648283600807, + 0.020578701049089432, + 0.010024560615420341, + 0.017745202407240868, + 0.010970224626362324, + 0.012815219350159168, + 0.014840241521596909, + 0.009823822416365147, + 0.007861695252358913, + 0.01048512477427721, + 0.007781567517668009, + 0.009471451863646507, + 0.006972134113311768, + 0.013560978695750237, + 0.0068709105253219604, + 0.012564483098685741, + 0.005570134148001671, + 0.010804497636854649, + 0.013940747827291489, + 0.006455425173044205, + 0.007198486477136612, + 0.02095497027039528, + 0.017203666269779205, + 0.021548857912421227, + 0.013249948620796204, + 0.014937479980289936, + 0.012134959921240807, + 0.01141966413706541, + 0.008794700726866722, + 0.0057460893876850605, + 0.004571054596453905, + 0.0074269408360123634, + 0.005545143969357014, + 0.004704486578702927, + 0.0050095864571630955, + 0.005223291460424662, + 0.008652116172015667, + 0.004548633471131325, + 0.004646273795515299, + 0.005261339712888002, + 0.005794792901724577, + 0.005431537516415119, + 0.005093133542686701, + 0.004843461327254772, + 0.005620899610221386, + 0.006492580287158489, + 0.009201419539749622, + 0.008728565648198128, + 0.008160875178873539, + 0.007765776477754116, + 0.005725288763642311, + 0.010624654591083527, + 0.007147330325096846, + 0.0023622242733836174, + 0.006526780780404806, + 0.004613571800291538, + 0.004125799983739853, + 0.004675619769841433, + 0.0025234974455088377, + 0.003671499202027917, + 0.003141234628856182, + 0.012303601019084454, + 0.020891098305583, + 0.006897804792970419, + 0.011000925675034523, + 0.028919100761413574, + 0.01082625798881054, + 0.012883777730166912, + 0.007622117176651955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0790054127573967 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062075622379779816 + ], + "opposite_logprob_tokens": [ + -2.812075614929199 + ], + "logprob_reward": -0.062075622379779816, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062075622379779816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23210126161575317, + 0.0802169144153595, + 0.04399004578590393, + 0.15583419799804688, + 0.04497257620096207, + 0.05177690461277962, + 0.07869703322649002, + 0.08310019969940186, + 0.05126266926527023, + 0.1780482679605484 + ], + "credit_map": [ + 0.02652573585510254, + 0.05405354127287865, + 0.04268317669630051, + 0.02026400715112686, + 0.0427657887339592, + 0.011338960379362106, + 0.019662905484437943, + 0.014807140454649925, + 0.010091379284858704, + 0.014108547009527683, + 0.0118458466604352, + 0.011426054872572422, + 0.011760422959923744, + 0.004954878240823746, + 0.0109921395778656, + 0.005037645809352398, + 0.006565474439412355, + 0.00929947104305029, + 0.006705430336296558, + 0.004394195508211851, + 0.00510590011253953, + 0.004613540135324001, + 0.003747455542907119, + 0.0035585782025009394, + 0.02191142365336418, + 0.020031943917274475, + 0.03240145742893219, + 0.017964255064725876, + 0.02033509686589241, + 0.016666630282998085, + 0.014188816770911217, + 0.012334568426012993, + 0.004320253152400255, + 0.00658767856657505, + 0.006092215422540903, + 0.008049462921917439, + 0.006197301670908928, + 0.003638175781816244, + 0.006088786758482456, + 0.003998700994998217, + 0.006398242898285389, + 0.0051037403754889965, + 0.008074610494077206, + 0.006639187224209309, + 0.004373652394860983, + 0.0054214121773839, + 0.007001745980232954, + 0.008764314465224743, + 0.007330076768994331, + 0.010819105431437492, + 0.010919181630015373, + 0.013202697969973087, + 0.00953590776771307, + 0.009166385978460312, + 0.009738261811435223, + 0.007985417731106281, + 0.007371632847934961, + 0.009353825822472572, + 0.013896682299673557, + 0.01025497168302536, + 0.011306947097182274, + 0.009323565289378166, + 0.01318947784602642, + 0.00840309914201498, + 0.003658047178760171, + 0.010859940201044083, + 0.006233662832528353, + 0.006338894367218018, + 0.006821284536272287, + 0.0032267356291413307, + 0.01002444140613079, + 0.004099664278328419, + 0.014199460856616497, + 0.02211308293044567, + 0.017681900411844254, + 0.03766626864671707, + 0.028988296166062355, + 0.01102894451469183, + 0.03014376573264599, + 0.016226543113589287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.062075622379779816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003197202691808343 + ], + "opposite_logprob_tokens": [ + -5.753197193145752 + ], + "logprob_reward": -0.003197202691808343, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003197202691808343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16360005736351013, + 0.10741794854402542, + 0.04094667360186577, + 0.13999038934707642, + 0.0516984723508358, + 0.04239905998110771, + 0.0826568454504013, + 0.05260487273335457, + 0.047416284680366516, + 0.27126941084861755 + ], + "credit_map": [ + 0.0175725556910038, + 0.03631075099110603, + 0.027931269258260727, + 0.012457071803510189, + 0.037704695016145706, + 0.009062597528100014, + 0.014420002698898315, + 0.008141119964420795, + 0.012504525482654572, + 0.014768846333026886, + 0.01847093179821968, + 0.014453482814133167, + 0.016612688079476357, + 0.010241660289466381, + 0.011759704910218716, + 0.008606120944023132, + 0.007136261090636253, + 0.009636937640607357, + 0.005141796078532934, + 0.004626761190593243, + 0.004819486290216446, + 0.0026614174712449312, + 0.0031063619535416365, + 0.0038176525849848986, + 0.02252035215497017, + 0.01526255626231432, + 0.020434634760022163, + 0.022315125912427902, + 0.01692315749824047, + 0.01374944020062685, + 0.017160193994641304, + 0.011624917387962341, + 0.007066919002681971, + 0.0067754024639725685, + 0.007803591433912516, + 0.008921550586819649, + 0.0061651659198105335, + 0.005201813764870167, + 0.005756959319114685, + 0.004007069393992424, + 0.005685916170477867, + 0.003970305435359478, + 0.005390609614551067, + 0.005841999314725399, + 0.0040927850641310215, + 0.005492778494954109, + 0.005421831738203764, + 0.006502833683043718, + 0.0073876879177987576, + 0.009108312427997589, + 0.0113446693867445, + 0.012639835476875305, + 0.011853602714836597, + 0.010123754851520061, + 0.010914026759564877, + 0.009284956380724907, + 0.00546885235235095, + 0.006792353931814432, + 0.00643492117524147, + 0.00700031965970993, + 0.007690494880080223, + 0.006273131351917982, + 0.007360541261732578, + 0.005584262311458588, + 0.004196725320070982, + 0.009523669257760048, + 0.004590660333633423, + 0.005681117996573448, + 0.007203665096312761, + 0.0034353246446698904, + 0.008017873391509056, + 0.004767248407006264, + 0.019707346335053444, + 0.02913501113653183, + 0.03030286729335785, + 0.0638837069272995, + 0.036340951919555664, + 0.015539168380200863, + 0.05843256413936615, + 0.017927786335349083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.003197202691808343 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007651190273463726 + ], + "opposite_logprob_tokens": [ + -4.882651329040527 + ], + "logprob_reward": -0.007651190273463726, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007651190273463726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32082366943359375, + 0.09286512434482574, + 0.04360802099108696, + 0.10117464512586594, + 0.09098226577043533, + 0.0477498322725296, + 0.12996666133403778, + 0.03842473030090332, + 0.04402979090809822, + 0.09037532657384872 + ], + "credit_map": [ + 0.043430525809526443, + 0.05720232054591179, + 0.060978446155786514, + 0.026368511840701103, + 0.06912462413311005, + 0.02119903825223446, + 0.02520926110446453, + 0.01731092855334282, + 0.014300752431154251, + 0.012743251398205757, + 0.012931662611663342, + 0.011918608099222183, + 0.01452314667403698, + 0.0065785921178758144, + 0.01205907016992569, + 0.007810031529515982, + 0.007285099010914564, + 0.008065703324973583, + 0.005608539562672377, + 0.004078140947967768, + 0.0052168467082083225, + 0.004747270606458187, + 0.00470587145537138, + 0.003900550538673997, + 0.018084431067109108, + 0.010329009965062141, + 0.016749808564782143, + 0.007894582115113735, + 0.011740743182599545, + 0.018405849114060402, + 0.008294891566038132, + 0.009675320237874985, + 0.018251046538352966, + 0.010765171609818935, + 0.0174275990575552, + 0.009209811687469482, + 0.009979535825550556, + 0.011425584554672241, + 0.0073883249424397945, + 0.006535183638334274, + 0.007009599823504686, + 0.006366061978042126, + 0.007012888789176941, + 0.007442311383783817, + 0.005770151969045401, + 0.0045301299542188644, + 0.004774671513587236, + 0.0048440126702189445, + 0.016841432079672813, + 0.027117634192109108, + 0.01546365674585104, + 0.010785449296236038, + 0.01755739562213421, + 0.015700841322541237, + 0.00911958608776331, + 0.01738065853714943, + 0.00441603222861886, + 0.008038598112761974, + 0.005205901805311441, + 0.0042168800719082355, + 0.005170919466763735, + 0.0033726654946804047, + 0.005046213045716286, + 0.0029575200751423836, + 0.003663846757262945, + 0.01253192126750946, + 0.00417893473058939, + 0.0034247059375047684, + 0.008579338900744915, + 0.00397974718362093, + 0.004146919120103121, + 0.0035243744496256113, + 0.009782429784536362, + 0.016410935670137405, + 0.005461887456476688, + 0.00902425218373537, + 0.02212822251021862, + 0.009975752793252468, + 0.010230638086795807, + 0.0073611983098089695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.007651190273463726 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up frames a crystal-clear cylindrical glass filled with perfectly still, transparent water, set against a dark, minimalist backdrop. At the bottom of the glass, a solid flake of white soapstone begins to dissolve violently. Instantly, a cloud of giant silver bubbles erupts from the entire surface, forming a shimmering vertical stream that rapidly rises through the water. As this continuous stream of bubbles reaches the surface, they accumulate and coalesce, meticulously forming an expanding layer of dense, pure white foam. The camera captures the dynamic contrast between the chaotic upward surge of bubbles and the steady, deliberate growth of a creamy layer of foam, which gradually thickens and covers the entire surface of the water. All illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the pristine texture of the foam, creating a focused and captivating atmosphere of vibrant transformation.", + "seed": 28419, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_4/index_6499_seed_28419_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.053678900562226774, + "vlm_logprob": -0.053678900562226774, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.053678900562226774, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02690081298351288 + ], + "opposite_logprob_tokens": [ + -5.651900768280029 + ], + "logprob_reward": -0.02690081298351288, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.02690081298351288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10124360024929047, + 0.06718257814645767, + 0.04554355517029762, + 0.06561176478862762, + 0.047849930822849274, + 0.04031076654791832, + 0.06175795942544937, + 0.06301235407590866, + 0.08129774034023285, + 0.4261897802352905 + ], + "credit_map": [ + 0.010620662942528725, + 0.017216110602021217, + 0.01353068184107542, + 0.009874118492007256, + 0.01921265199780464, + 0.009170370176434517, + 0.009900015778839588, + 0.011718985624611378, + 0.00874157715588808, + 0.00920074712485075, + 0.006728029809892178, + 0.010857926681637764, + 0.008358582854270935, + 0.008506784215569496, + 0.008308511227369308, + 0.006480411160737276, + 0.005860976874828339, + 0.005154133308678865, + 0.005795391742140055, + 0.004936175420880318, + 0.004222119227051735, + 0.007738759275525808, + 0.005133436527103186, + 0.0067025586031377316, + 0.005810075905174017, + 0.010875113308429718, + 0.007044778671115637, + 0.00656760623678565, + 0.009331912733614445, + 0.007129153702408075, + 0.01209526602178812, + 0.006757854018360376, + 0.0062764850445091724, + 0.005966339726001024, + 0.007147625088691711, + 0.006050503347069025, + 0.004929760005325079, + 0.00654810993000865, + 0.005629695951938629, + 0.005301411729305983, + 0.005289615131914616, + 0.004474477376788855, + 0.006060727406293154, + 0.005119674373418093, + 0.004544349852949381, + 0.00447422219440341, + 0.005412515718489885, + 0.0049351793713867664, + 0.00431973859667778, + 0.006070361472666264, + 0.00538983941078186, + 0.0121065154671669, + 0.004563909024000168, + 0.003387525910511613, + 0.01789472997188568, + 0.0080253379419446, + 0.0066621690057218075, + 0.010634295642375946, + 0.010040252469480038, + 0.007575559429824352, + 0.006771124433726072, + 0.005363638047128916, + 0.01047718059271574, + 0.005488126538693905, + 0.005195084027945995, + 0.012796946801245213, + 0.018166838213801384, + 0.00955174956470728, + 0.01815088465809822, + 0.006427663378417492, + 0.005530572962015867, + 0.0054779923520982265, + 0.0147714177146554, + 0.16685791313648224, + 0.03969862312078476, + 0.025265026837587357, + 0.14481337368488312, + 0.012946098111569881, + 0.013729353435337543, + 0.008107936009764671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.02690081298351288 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982296794652939 + ], + "opposite_logprob_tokens": [ + -3.529823064804077 + ], + "logprob_reward": -0.02982296794652939, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982296794652939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1801711618900299, + 0.10151118040084839, + 0.053511280566453934, + 0.0673360824584961, + 0.12140782177448273, + 0.053236011415719986, + 0.046501677483320236, + 0.035755760967731476, + 0.09404678642749786, + 0.2465222030878067 + ], + "credit_map": [ + 0.019078440964221954, + 0.027945689857006073, + 0.020211542025208473, + 0.029799934476614, + 0.02514083869755268, + 0.018035538494586945, + 0.020469365641474724, + 0.019489819183945656, + 0.011345128528773785, + 0.012398771941661835, + 0.010196813382208347, + 0.012043996714055538, + 0.00920040998607874, + 0.021911486983299255, + 0.00936050247400999, + 0.015054073184728622, + 0.006108333822339773, + 0.006364355329424143, + 0.00591876357793808, + 0.005514063406735659, + 0.006963803432881832, + 0.007984363473951817, + 0.008092056959867477, + 0.006565543822944164, + 0.01003856398165226, + 0.00714488048106432, + 0.0068537406623363495, + 0.007909051142632961, + 0.007783378940075636, + 0.0111099136993289, + 0.007658636197447777, + 0.008837924338877201, + 0.013829536736011505, + 0.014813754707574844, + 0.019530881196260452, + 0.014910716563463211, + 0.01203614380210638, + 0.018301410600543022, + 0.014351475983858109, + 0.01363392360508442, + 0.010913588106632233, + 0.005759605206549168, + 0.005720358807593584, + 0.005939323455095291, + 0.005610557273030281, + 0.005202359985560179, + 0.0072977677918970585, + 0.006792452186346054, + 0.004085631109774113, + 0.005246632732450962, + 0.004674443509429693, + 0.007813735865056515, + 0.004321391228586435, + 0.0032634015660732985, + 0.010422632098197937, + 0.006673810537904501, + 0.003383710514754057, + 0.00540906423702836, + 0.004698186181485653, + 0.004904739558696747, + 0.004133476410061121, + 0.0030399663373827934, + 0.005215183831751347, + 0.004971434362232685, + 0.007488928735256195, + 0.012818082235753536, + 0.021588753908872604, + 0.013057212345302105, + 0.015065968967974186, + 0.0077806017361581326, + 0.010274660773575306, + 0.005972588900476694, + 0.011110207065939903, + 0.04395361244678497, + 0.06393854320049286, + 0.030277539044618607, + 0.06916715949773788, + 0.009582942351698875, + 0.011960783042013645, + 0.006531453691422939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.02982296794652939 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20160385966300964 + ], + "opposite_logprob_tokens": [ + -1.701603889465332 + ], + "logprob_reward": -0.20160385966300964, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160385966300964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10442014783620834, + 0.09867473691701889, + 0.13568376004695892, + 0.053475864231586456, + 0.05810749903321266, + 0.079087033867836, + 0.08163364231586456, + 0.09793558716773987, + 0.05999519303441048, + 0.23098652064800262 + ], + "credit_map": [ + 0.009924102574586868, + 0.015031000599265099, + 0.01749793253839016, + 0.013457337394356728, + 0.012950601056218147, + 0.013909105211496353, + 0.00987610500305891, + 0.011773954145610332, + 0.009774407371878624, + 0.015363655984401703, + 0.00904550775885582, + 0.022375551983714104, + 0.01271780114620924, + 0.013113793917000294, + 0.007171226665377617, + 0.009112786501646042, + 0.014367391355335712, + 0.017455069348216057, + 0.015295242890715599, + 0.013992422260344028, + 0.016690025106072426, + 0.019103409722447395, + 0.018462546169757843, + 0.02031763456761837, + 0.005093082319945097, + 0.0068947309628129005, + 0.006933970842510462, + 0.0065475692972540855, + 0.004690614063292742, + 0.008512181229889393, + 0.00669354060664773, + 0.0081101693212986, + 0.0052580260671675205, + 0.00843366514891386, + 0.005940642207860947, + 0.005951251834630966, + 0.006956610828638077, + 0.009367795661091805, + 0.006780051160603762, + 0.009419449605047703, + 0.007996691390872002, + 0.008976244367659092, + 0.011589308269321918, + 0.012025094591081142, + 0.006141494493931532, + 0.00814798567444086, + 0.014086307026445866, + 0.010123897343873978, + 0.003838157746940851, + 0.006066563073545694, + 0.007189343683421612, + 0.0191205944865942, + 0.005182073451578617, + 0.003117714310064912, + 0.02923274226486683, + 0.007886441424489021, + 0.009152975864708424, + 0.015206810086965561, + 0.017110880464315414, + 0.011646597646176815, + 0.01621099002659321, + 0.008616114035248756, + 0.01228373497724533, + 0.007707472890615463, + 0.006259981542825699, + 0.011777195148169994, + 0.004620791878551245, + 0.006190677639096975, + 0.012648472562432289, + 0.004565712064504623, + 0.008203644305467606, + 0.005728711374104023, + 0.008549736812710762, + 0.08284081518650055, + 0.03634103015065193, + 0.019007453694939613, + 0.06653100252151489, + 0.007120211608707905, + 0.00707244873046875, + 0.0035238037817180157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.20160385966300964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954621359705925 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954621359705925, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954621359705925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13401158154010773, + 0.05281009525060654, + 0.05834059789776802, + 0.05505795031785965, + 0.032241642475128174, + 0.08023978769779205, + 0.10082121938467026, + 0.09851528704166412, + 0.2512509822845459, + 0.13671083748340607 + ], + "credit_map": [ + 0.010843990370631218, + 0.021168557927012444, + 0.022453954443335533, + 0.016816211864352226, + 0.01784767396748066, + 0.014805655926465988, + 0.016757095232605934, + 0.013318446464836597, + 0.007142568472772837, + 0.009171925485134125, + 0.005441426299512386, + 0.006311290431767702, + 0.007242989726364613, + 0.006615304388105869, + 0.006197499576956034, + 0.004687086679041386, + 0.0059500024653971195, + 0.010953648947179317, + 0.006055628415197134, + 0.006138488184660673, + 0.008811798878014088, + 0.007961232215166092, + 0.006078425794839859, + 0.006391372997313738, + 0.005949812475591898, + 0.008288778364658356, + 0.0071848006919026375, + 0.006091005634516478, + 0.006837944965809584, + 0.0064691505394876, + 0.0076999356970191, + 0.006536523345857859, + 0.005101716611534357, + 0.003101779380813241, + 0.003999608568847179, + 0.003492954885587096, + 0.004177538678050041, + 0.004389808047562838, + 0.002941808896139264, + 0.005036426242440939, + 0.008975744247436523, + 0.00872194766998291, + 0.010349988006055355, + 0.011566843837499619, + 0.007558802142739296, + 0.008654305711388588, + 0.013504553586244583, + 0.01090761087834835, + 0.0054587326012551785, + 0.008320351131260395, + 0.011054215021431446, + 0.021374166011810303, + 0.009028036147356033, + 0.005451009143143892, + 0.030364755541086197, + 0.009769956581294537, + 0.008271812461316586, + 0.017867395654320717, + 0.010111293755471706, + 0.015537416562438011, + 0.010865104384720325, + 0.009006510488688946, + 0.017428269609808922, + 0.00942748598754406, + 0.017146706581115723, + 0.04030708968639374, + 0.04803931713104248, + 0.030808497220277786, + 0.05687907710671425, + 0.0184489656239748, + 0.02546398714184761, + 0.01415734738111496, + 0.007045122794806957, + 0.03923031687736511, + 0.03292010352015495, + 0.01692911423742771, + 0.022769467905163765, + 0.006228748708963394, + 0.007488830015063286, + 0.004099119454622269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.005954621359705925 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004112240858376026 + ], + "opposite_logprob_tokens": [ + -5.504112243652344 + ], + "logprob_reward": -0.004112240858376026, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004112240858376026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2153845876455307, + 0.05088294297456741, + 0.042723577469587326, + 0.044887714087963104, + 0.04903070256114006, + 0.03684878721833229, + 0.043910153210163116, + 0.09193956851959229, + 0.11104961484670639, + 0.31334245204925537 + ], + "credit_map": [ + 0.01869054324924946, + 0.03304377570748329, + 0.03843832015991211, + 0.024241486564278603, + 0.03293953090906143, + 0.02465960383415222, + 0.025810163468122482, + 0.017561133950948715, + 0.006187376566231251, + 0.006363390479236841, + 0.005182536784559488, + 0.006298222113400698, + 0.0062034581787884235, + 0.0080873416736722, + 0.0062749735079705715, + 0.006285637151449919, + 0.0047219255939126015, + 0.007151497527956963, + 0.004744802601635456, + 0.004578176885843277, + 0.00566220423206687, + 0.006206974387168884, + 0.005316104739904404, + 0.0043418873101472855, + 0.005438065156340599, + 0.004765554331243038, + 0.004456297028809786, + 0.005012675654143095, + 0.004976434633135796, + 0.006824403535574675, + 0.007036070339381695, + 0.006378208752721548, + 0.006320627871900797, + 0.006807594560086727, + 0.006145415361970663, + 0.005619882605969906, + 0.0049504744820296764, + 0.007769858930259943, + 0.005712559446692467, + 0.005704281851649284, + 0.0038050683215260506, + 0.003871731460094452, + 0.004250857047736645, + 0.005075489170849323, + 0.004347565583884716, + 0.0046472870744764805, + 0.00474945455789566, + 0.006101331673562527, + 0.0031004506163299084, + 0.004899072460830212, + 0.0039719645865261555, + 0.006938121281564236, + 0.004824427422136068, + 0.003240063088014722, + 0.009915736503899097, + 0.007020311895757914, + 0.008229144848883152, + 0.01423585508018732, + 0.009468364529311657, + 0.013705985620617867, + 0.010879818350076675, + 0.010096918791532516, + 0.017067359760403633, + 0.008256113156676292, + 0.009112665429711342, + 0.013765361160039902, + 0.02958807907998562, + 0.01519998349249363, + 0.01775633916258812, + 0.00881345197558403, + 0.008893229067325592, + 0.007920491509139538, + 0.018014322966337204, + 0.07202986627817154, + 0.1032489538192749, + 0.050165239721536636, + 0.029235368594527245, + 0.014034554362297058, + 0.01766319014132023, + 0.008950937539339066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.004112240858376026 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up frames a crystal-clear cylindrical glass filled with perfectly still, transparent water, set against a dark, minimalist backdrop. At the bottom of the glass, a solid flake of white soapstone begins to dissolve violently. Instantly, a cloud of giant silver bubbles erupts from the entire surface, forming a shimmering vertical stream that rapidly rises through the water. As this continuous stream of bubbles reaches the surface, they accumulate and coalesce, meticulously forming an expanding layer of dense, pure white foam. The camera captures the dynamic contrast between the chaotic upward surge of bubbles and the steady, deliberate growth of a creamy layer of foam, which gradually thickens and covers the entire surface of the water. All illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the pristine texture of the foam, creating a focused and captivating atmosphere of vibrant transformation.", + "seed": 28420, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_4/index_6499_seed_28420_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02867184178903699, + "vlm_logprob": -0.02867184178903699, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.02867184178903699, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0185096375644207 + ], + "opposite_logprob_tokens": [ + -5.143509864807129 + ], + "logprob_reward": -0.0185096375644207, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0185096375644207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05717618390917778, + 0.04818761721253395, + 0.08281013369560242, + 0.05967840552330017, + 0.04018010199069977, + 0.06439568847417831, + 0.13010452687740326, + 0.1632942408323288, + 0.27564528584480286, + 0.07852772623300552 + ], + "credit_map": [ + 0.007724015507847071, + 0.007886937819421291, + 0.006365835666656494, + 0.009331434965133667, + 0.005394641309976578, + 0.007457266096025705, + 0.007541488390415907, + 0.005474566947668791, + 0.007280939258635044, + 0.006080066319555044, + 0.004793152678757906, + 0.006152728572487831, + 0.00510649848729372, + 0.0053865849040448666, + 0.008588125929236412, + 0.004799525719136, + 0.0033787034917622805, + 0.004775972105562687, + 0.004728918429464102, + 0.005382954608649015, + 0.004648580215871334, + 0.023225784301757812, + 0.030835486948490143, + 0.005833737552165985, + 0.005451985634863377, + 0.01355065032839775, + 0.005270084366202354, + 0.006162009201943874, + 0.0061312345787882805, + 0.00750484736636281, + 0.00872483104467392, + 0.0068827723152935505, + 0.0032768920063972473, + 0.003950712736696005, + 0.005022140685468912, + 0.0034281094558537006, + 0.003954654093831778, + 0.009162298403680325, + 0.004421246703714132, + 0.0069640520960092545, + 0.00975947268307209, + 0.010164763778448105, + 0.009640248492360115, + 0.0072252326644957066, + 0.004257713444530964, + 0.007725049275904894, + 0.006313965190201998, + 0.00930925365537405, + 0.011660201475024223, + 0.020429564639925957, + 0.017403673380613327, + 0.012458257377147675, + 0.021676432341337204, + 0.015258259139955044, + 0.016417041420936584, + 0.014801113866269588, + 0.011459970846772194, + 0.043205853551626205, + 0.009361534379422665, + 0.017210083082318306, + 0.03085966594517231, + 0.01845024898648262, + 0.017432130873203278, + 0.015314762480556965, + 0.02220694161951542, + 0.03894652798771858, + 0.014908802695572376, + 0.07446515560150146, + 0.034810617566108704, + 0.031101811677217484, + 0.03681451082229614, + 0.02239094115793705, + 0.007580308709293604, + 0.011865993961691856, + 0.01048921886831522, + 0.00968520250171423, + 0.013219499960541725, + 0.010271569713950157, + 0.009482242166996002, + 0.005933696404099464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.0185096375644207 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811933100223541 + ], + "opposite_logprob_tokens": [ + -3.288119316101074 + ], + "logprob_reward": -0.03811933100223541, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811933100223541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21105091273784637, + 0.045285020023584366, + 0.22495965659618378, + 0.10037737339735031, + 0.05078738555312157, + 0.06448476016521454, + 0.052125584334135056, + 0.04722866415977478, + 0.12538796663284302, + 0.07831266522407532 + ], + "credit_map": [ + 0.03054114617407322, + 0.03400325030088425, + 0.016645900905132294, + 0.04226023703813553, + 0.02665811777114868, + 0.023678520694375038, + 0.019666627049446106, + 0.017597107216715813, + 0.005300601944327354, + 0.003936230204999447, + 0.005315593909472227, + 0.006189438048750162, + 0.0046685077250003815, + 0.004794232547283173, + 0.008663191460072994, + 0.006417223252356052, + 0.00901245791465044, + 0.009407908655703068, + 0.009402647614479065, + 0.01372706051915884, + 0.01378906425088644, + 0.0703829899430275, + 0.08538266271352768, + 0.013854856602847576, + 0.012625575996935368, + 0.02550249546766281, + 0.007799904327839613, + 0.011713298037648201, + 0.010063163936138153, + 0.011423593387007713, + 0.012421280145645142, + 0.008828061632812023, + 0.004046671558171511, + 0.005913793109357357, + 0.007887450978159904, + 0.004323654342442751, + 0.007519809063524008, + 0.008931640535593033, + 0.005178001243621111, + 0.006986364722251892, + 0.008084245957434177, + 0.008052543736994267, + 0.010012057609856129, + 0.008633737452328205, + 0.007116509135812521, + 0.007177641615271568, + 0.005918619688600302, + 0.009489406831562519, + 0.004596531856805086, + 0.008368316106498241, + 0.0067032091319561005, + 0.005143765360116959, + 0.008668769150972366, + 0.006088987924158573, + 0.006075295619666576, + 0.006480703596025705, + 0.004607343580573797, + 0.010551071725785732, + 0.003573895199224353, + 0.004508601501584053, + 0.007032726425677538, + 0.00794938299804926, + 0.004655462224036455, + 0.004350181203335524, + 0.010570472106337547, + 0.02551850490272045, + 0.00796218030154705, + 0.011850296519696712, + 0.025859184563159943, + 0.019515320658683777, + 0.015686804428696632, + 0.008425208739936352, + 0.007400874514132738, + 0.010877201333642006, + 0.008914243429899216, + 0.010559119284152985, + 0.012513873167335987, + 0.011326312087476254, + 0.010343263857066631, + 0.006377776153385639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.03811933100223541 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900188863277435 + ], + "opposite_logprob_tokens": [ + -2.5790019035339355 + ], + "logprob_reward": -0.07900188863277435, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900188863277435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13251033425331116, + 0.0443454347550869, + 0.03911975398659706, + 0.2288987934589386, + 0.10739348083734512, + 0.09329316765069962, + 0.10719158500432968, + 0.08563796430826187, + 0.038612958043813705, + 0.12299655377864838 + ], + "credit_map": [ + 0.017166713252663612, + 0.018251430243253708, + 0.011673593893647194, + 0.021274589002132416, + 0.018954236060380936, + 0.01385636068880558, + 0.01891420967876911, + 0.012419198639690876, + 0.005900786258280277, + 0.005392776802182198, + 0.004495585802942514, + 0.006130371708422899, + 0.006812255829572678, + 0.004288427531719208, + 0.00673760985955596, + 0.004587620496749878, + 0.002974150003865361, + 0.004696459509432316, + 0.004838436376303434, + 0.003545813960954547, + 0.004191215615719557, + 0.006622775923460722, + 0.006299648433923721, + 0.005951250903308392, + 0.020647753030061722, + 0.054927051067352295, + 0.019646858796477318, + 0.02346031181514263, + 0.026206443086266518, + 0.029223527759313583, + 0.03145147114992142, + 0.023335367441177368, + 0.008532729931175709, + 0.01195601187646389, + 0.014826011843979359, + 0.013299867510795593, + 0.012225930579006672, + 0.015746433287858963, + 0.0170070119202137, + 0.013799494132399559, + 0.01228349469602108, + 0.012433361262083054, + 0.014965897426009178, + 0.012923220172524452, + 0.008490148931741714, + 0.00874669011682272, + 0.011779245920479298, + 0.011671102605760098, + 0.009385921992361546, + 0.014300589449703693, + 0.012519648298621178, + 0.012278556823730469, + 0.016724741086363792, + 0.015564389526844025, + 0.013698114082217216, + 0.01271961722522974, + 0.006377895828336477, + 0.021182389929890633, + 0.006056325975805521, + 0.007828734815120697, + 0.013113928958773613, + 0.011816668324172497, + 0.00927454698830843, + 0.009987475350499153, + 0.0028805509209632874, + 0.005759993102401495, + 0.002969243098050356, + 0.007524300832301378, + 0.004994528833776712, + 0.006991473492234945, + 0.004701042547821999, + 0.0027918252162635326, + 0.011648020707070827, + 0.018151842057704926, + 0.011175120249390602, + 0.017203861847519875, + 0.021444858983159065, + 0.016084937378764153, + 0.017409464344382286, + 0.00987844355404377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.07900188863277435 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100131336599588 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.004100131336599588, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100131336599588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10229827463626862, + 0.05680966377258301, + 0.13095195591449738, + 0.20912228524684906, + 0.08315912634134293, + 0.03814011812210083, + 0.08902830630540848, + 0.07496292144060135, + 0.0917740911245346, + 0.12375328689813614 + ], + "credit_map": [ + 0.013712799176573753, + 0.014396488666534424, + 0.00958310067653656, + 0.017272107303142548, + 0.012645457871258259, + 0.011003500781953335, + 0.013131873682141304, + 0.010552948340773582, + 0.0070609585382044315, + 0.005327986087650061, + 0.004643453750759363, + 0.00790550746023655, + 0.006682797335088253, + 0.006192984990775585, + 0.012040562927722931, + 0.00695541687309742, + 0.0054790726862847805, + 0.005563050042837858, + 0.006728982087224722, + 0.007227674126625061, + 0.007779515814036131, + 0.04148352891206741, + 0.04800921306014061, + 0.008680925704538822, + 0.019797449931502342, + 0.05013439431786537, + 0.016863012686371803, + 0.021377939730882645, + 0.025553761050105095, + 0.02647482231259346, + 0.026556305587291718, + 0.02236458659172058, + 0.009419556707143784, + 0.008530860766768456, + 0.016271894797682762, + 0.007038122974336147, + 0.00914483331143856, + 0.009888550266623497, + 0.01087270863354206, + 0.011992600746452808, + 0.006896755192428827, + 0.004509874153882265, + 0.004861840978264809, + 0.003823766252025962, + 0.003631607396528125, + 0.005162475630640984, + 0.004266683477908373, + 0.004987115040421486, + 0.008321264758706093, + 0.012322463095188141, + 0.009431962855160236, + 0.01244142185896635, + 0.013691885396838188, + 0.01172900851815939, + 0.009611277841031551, + 0.011479020118713379, + 0.006011556833982468, + 0.018959764391183853, + 0.005201936233788729, + 0.007616683375090361, + 0.012259325943887234, + 0.009868265129625797, + 0.0073005566373467445, + 0.0077448352240026, + 0.007275239564478397, + 0.016296425834298134, + 0.007891345769166946, + 0.010419891215860844, + 0.013375218957662582, + 0.01637566089630127, + 0.013089657761156559, + 0.007050651591271162, + 0.011354510672390461, + 0.014006991870701313, + 0.012491971254348755, + 0.018624197691679, + 0.016273541375994682, + 0.016741426661610603, + 0.019314544275403023, + 0.014946102164685726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.004100131336599588 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003628220409154892 + ], + "opposite_logprob_tokens": [ + -5.628628253936768 + ], + "logprob_reward": -0.003628220409154892, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003628220409154892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13150636851787567, + 0.057177577167749405, + 0.25523480772972107, + 0.15843559801578522, + 0.057056620717048645, + 0.03632744774222374, + 0.03181085363030434, + 0.06801631301641464, + 0.0802004262804985, + 0.12423401325941086 + ], + "credit_map": [ + 0.01728811487555504, + 0.02264929935336113, + 0.008608321659266949, + 0.03131647780537605, + 0.014930379576981068, + 0.01719343662261963, + 0.01138748787343502, + 0.008132850751280785, + 0.006836733780801296, + 0.005423489958047867, + 0.004503215197473764, + 0.010378001257777214, + 0.007342703640460968, + 0.006125127896666527, + 0.008599885739386082, + 0.00796841736882925, + 0.01501236017793417, + 0.015655891969799995, + 0.014954962767660618, + 0.015581944026052952, + 0.01707332767546177, + 0.07169827818870544, + 0.0857158750295639, + 0.019542166963219643, + 0.01696491241455078, + 0.04075905680656433, + 0.013023902662098408, + 0.017352616414427757, + 0.01591307297348976, + 0.019573096185922623, + 0.018723737448453903, + 0.016125204041600227, + 0.006415335927158594, + 0.0067304158583283424, + 0.009752143174409866, + 0.0046646068803966045, + 0.005576750263571739, + 0.008131807669997215, + 0.0071663446724414825, + 0.0086192162707448, + 0.004690902773290873, + 0.004952665884047747, + 0.005923592485487461, + 0.0038719158619642258, + 0.003652355633676052, + 0.004487624857574701, + 0.003900590119883418, + 0.0048477970995008945, + 0.003327394602820277, + 0.00443511176854372, + 0.003909631632268429, + 0.0033582462929189205, + 0.005962996277958155, + 0.004703073296695948, + 0.002775573404505849, + 0.0033388251904398203, + 0.004896864760667086, + 0.01997663639485836, + 0.005477373022586107, + 0.004841443616896868, + 0.010659546591341496, + 0.009163685142993927, + 0.004673881456255913, + 0.008326881565153599, + 0.006194727495312691, + 0.013654054142534733, + 0.005102748051285744, + 0.018927903845906258, + 0.012121371924877167, + 0.010473333299160004, + 0.008162353187799454, + 0.005563931539654732, + 0.01244365144520998, + 0.01753230392932892, + 0.013624078594148159, + 0.01790500059723854, + 0.017888886854052544, + 0.019869545474648476, + 0.013665378093719482, + 0.011305168271064758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.003628220409154892 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angle close-up frames a crystal-clear cylindrical glass filled with perfectly still, transparent water, set against a dark, minimalist backdrop. At the bottom of the glass, a solid flake of white soapstone begins to dissolve violently. Instantly, a cloud of giant silver bubbles erupts from the entire surface, forming a shimmering vertical stream that rapidly rises through the water. As this continuous stream of bubbles reaches the surface, they accumulate and coalesce, meticulously forming an expanding layer of dense, pure white foam. The camera captures the dynamic contrast between the chaotic upward surge of bubbles and the steady, deliberate growth of a creamy layer of foam, which gradually thickens and covers the entire surface of the water. All illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the pristine texture of the foam, creating a focused and captivating atmosphere of vibrant transformation.", + "seed": 28421, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_4/index_6499_seed_28421_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10486045179422945, + "vlm_logprob": -0.10486045179422945, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.10486045179422945, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015718374401330948 + ], + "opposite_logprob_tokens": [ + -6.515718460083008 + ], + "logprob_reward": -0.015718374401330948, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.015718374401330948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10626427829265594, + 0.20637434720993042, + 0.07250151038169861, + 0.09296299517154694, + 0.09452579915523529, + 0.09251067787408829, + 0.11456841230392456, + 0.07806822657585144, + 0.09271250665187836, + 0.04951128363609314 + ], + "credit_map": [ + 0.010809315368533134, + 0.01101559866219759, + 0.011015616357326508, + 0.017637021839618683, + 0.011293687857687473, + 0.008404377847909927, + 0.02383827231824398, + 0.012250389903783798, + 0.013722311705350876, + 0.015529480762779713, + 0.04975714907050133, + 0.06295745819807053, + 0.013056565076112747, + 0.00716846389696002, + 0.023594798520207405, + 0.020588114857673645, + 0.01188720390200615, + 0.007797189988195896, + 0.006257099099457264, + 0.010607590898871422, + 0.007601489778608084, + 0.006404508370906115, + 0.01200235728174448, + 0.009944071061909199, + 0.012203183956444263, + 0.012625307776033878, + 0.009034096263349056, + 0.010611717589199543, + 0.011317657306790352, + 0.011960810981690884, + 0.015561249107122421, + 0.00964896846562624, + 0.008864938281476498, + 0.01111735962331295, + 0.008170592598617077, + 0.015657249838113785, + 0.013308033347129822, + 0.010724460706114769, + 0.01617528311908245, + 0.010507877916097641, + 0.007387526798993349, + 0.008800270967185497, + 0.010163516737520695, + 0.014745095744729042, + 0.014912591315805912, + 0.011956625618040562, + 0.014566902071237564, + 0.009978153742849827, + 0.0077631655149161816, + 0.01207017432898283, + 0.014255412854254246, + 0.018221737816929817, + 0.013670826330780983, + 0.012792281806468964, + 0.021061211824417114, + 0.014733606949448586, + 0.0050057219341397285, + 0.011460353620350361, + 0.009684039279818535, + 0.01034032553434372, + 0.012852052226662636, + 0.009188163094222546, + 0.01027029287070036, + 0.009267279878258705, + 0.007755874656140804, + 0.020443487912416458, + 0.006881944835186005, + 0.010435185395181179, + 0.021720120683312416, + 0.009939332492649555, + 0.008799036964774132, + 0.006737519521266222, + 0.003767730202525854, + 0.011204882524907589, + 0.004040001425892115, + 0.0061764828860759735, + 0.008995890617370605, + 0.006129788234829903, + 0.004150740802288055, + 0.00504576601088047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.015718374401330948 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026362799108028412 + ], + "opposite_logprob_tokens": [ + -3.651362895965576 + ], + "logprob_reward": -0.026362799108028412, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026362799108028412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11742257326841354, + 0.03464210405945778, + 0.10139483213424683, + 0.16718702018260956, + 0.07744520157575607, + 0.12653711438179016, + 0.21040396392345428, + 0.056537993252277374, + 0.07745128870010376, + 0.03097795695066452 + ], + "credit_map": [ + 0.014401753433048725, + 0.010011625476181507, + 0.00940591562539339, + 0.016650410369038582, + 0.017016761004924774, + 0.016438912600278854, + 0.02216116338968277, + 0.011336038820445538, + 0.003586569568142295, + 0.0029543512500822544, + 0.003920041490346193, + 0.007634118665009737, + 0.003851457964628935, + 0.002673639450222254, + 0.006149915046989918, + 0.003872013185173273, + 0.011712709441781044, + 0.01105392724275589, + 0.011218546889722347, + 0.011624136939644814, + 0.010910624638199806, + 0.010786447674036026, + 0.016500765457749367, + 0.017587674781680107, + 0.029127372428774834, + 0.016808196902275085, + 0.01585301198065281, + 0.01946411095559597, + 0.027862006798386574, + 0.020033089444041252, + 0.020450714975595474, + 0.017588509246706963, + 0.01052520889788866, + 0.009319931268692017, + 0.008384833112359047, + 0.007950601167976856, + 0.01016139704734087, + 0.008625594899058342, + 0.00912952609360218, + 0.013348110020160675, + 0.010938423685729504, + 0.01412420067936182, + 0.013428745791316032, + 0.015443277545273304, + 0.026861486956477165, + 0.015304808504879475, + 0.016033155843615532, + 0.01440301351249218, + 0.014959615655243397, + 0.045132555067539215, + 0.021171389147639275, + 0.02742631919682026, + 0.030936965718865395, + 0.018693268299102783, + 0.034307949244976044, + 0.017775915563106537, + 0.003867242019623518, + 0.010383320041000843, + 0.00534448167309165, + 0.007029498927295208, + 0.011813191697001457, + 0.007220316678285599, + 0.006459057331085205, + 0.004420886747539043, + 0.005430418998003006, + 0.021167324855923653, + 0.0038440010976046324, + 0.005720123648643494, + 0.021850358694791794, + 0.007522780913859606, + 0.00576644204556942, + 0.0061498419381678104, + 0.002820426132529974, + 0.007507478352636099, + 0.0022798182908445597, + 0.003323434153571725, + 0.005254216026514769, + 0.00338591611944139, + 0.0022712296340614557, + 0.004135440103709698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.026362799108028412 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742966294288635 + ], + "opposite_logprob_tokens": [ + -0.9742966294288635 + ], + "logprob_reward": -0.4742966294288635, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742966294288635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028541911393404007, + 0.11436998099088669, + 0.07012198120355606, + 0.5931620001792908, + 0.06017936393618584, + 0.03200705721974373, + 0.035883624106645584, + 0.03458273410797119, + 0.01963796652853489, + 0.011513435281813145 + ], + "credit_map": [ + 0.0037541408091783524, + 0.0029300490859895945, + 0.00287949712947011, + 0.005672871600836515, + 0.003094895975664258, + 0.0023146027233451605, + 0.005127512849867344, + 0.002768340753391385, + 0.007176792714744806, + 0.008827692829072475, + 0.02948937378823757, + 0.03168407827615738, + 0.008931571617722511, + 0.0039744386449456215, + 0.011772424913942814, + 0.012513593770563602, + 0.007010277360677719, + 0.007204535882920027, + 0.007228866685181856, + 0.008505836129188538, + 0.0068562692031264305, + 0.007977155037224293, + 0.012508954852819443, + 0.012830079533159733, + 0.0628729984164238, + 0.06407792866230011, + 0.05914771184325218, + 0.07590410858392715, + 0.11793452501296997, + 0.06472205370664597, + 0.0915612205862999, + 0.056941378861665726, + 0.006922261789441109, + 0.008040648885071278, + 0.004221867769956589, + 0.007747425232082605, + 0.009776738472282887, + 0.006805973127484322, + 0.010259008966386318, + 0.006405435036867857, + 0.0024153892882168293, + 0.0043704393319785595, + 0.002988879568874836, + 0.004313057288527489, + 0.006911938078701496, + 0.0032849153503775597, + 0.0037637476343661547, + 0.003958684392273426, + 0.002567351795732975, + 0.004882031586021185, + 0.004101076163351536, + 0.004865885246545076, + 0.005881401710212231, + 0.003592289751395583, + 0.005659791175276041, + 0.004333794116973877, + 0.002443620702251792, + 0.005288260523229837, + 0.003252206137403846, + 0.00352833466604352, + 0.009801940061151981, + 0.004540807101875544, + 0.0031896312721073627, + 0.002537930617108941, + 0.001059574424289167, + 0.0034834728576242924, + 0.0013778809225186706, + 0.0031197876669466496, + 0.004564232658594847, + 0.0016656400403007865, + 0.0025299033150076866, + 0.0018374737119302154, + 0.0009079756564460695, + 0.004052390344440937, + 0.0004633896460290998, + 0.0008747863466851413, + 0.003082160372287035, + 0.0007819423917680979, + 0.0007521305815316737, + 0.0005986593314446509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.4742966294288635 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739859934896231 + ], + "opposite_logprob_tokens": [ + -5.006740093231201 + ], + "logprob_reward": -0.006739859934896231, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739859934896231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10360921919345856, + 0.1202610433101654, + 0.03806222230195999, + 0.2869364321231842, + 0.048470914363861084, + 0.04801797494292259, + 0.029989538714289665, + 0.18515168130397797, + 0.1255297064781189, + 0.013971279375255108 + ], + "credit_map": [ + 0.010830439627170563, + 0.012233712710440159, + 0.011099787428975105, + 0.016380641609430313, + 0.021185141056776047, + 0.008072278462350368, + 0.015623481944203377, + 0.008183734491467476, + 0.0074040633626282215, + 0.007313786074519157, + 0.03762869909405708, + 0.030416280031204224, + 0.005945518612861633, + 0.0030507370829582214, + 0.011859110556542873, + 0.016642844304442406, + 0.003916621673852205, + 0.004164636600762606, + 0.003638630034402013, + 0.005466359201818705, + 0.004085129592567682, + 0.004937722813338041, + 0.005716041661798954, + 0.006137080956250429, + 0.02543492242693901, + 0.03006686270236969, + 0.027835002169013023, + 0.0387459434568882, + 0.0587298646569252, + 0.031226303428411484, + 0.04884288087487221, + 0.026054631918668747, + 0.005053795408457518, + 0.00696934200823307, + 0.004218201152980328, + 0.0067771137692034245, + 0.0067728557623922825, + 0.005284569691866636, + 0.007992053404450417, + 0.005402983631938696, + 0.0036774713080376387, + 0.006938050035387278, + 0.003951709251850843, + 0.006498591508716345, + 0.013538549654185772, + 0.004662629216909409, + 0.0055860355496406555, + 0.0031649398151785135, + 0.0022154683247208595, + 0.00399246159940958, + 0.002773699816316366, + 0.0041644033044576645, + 0.006346045061945915, + 0.0037542094942182302, + 0.004030682146549225, + 0.0027125703636556864, + 0.012140228413045406, + 0.030730968341231346, + 0.016927653923630714, + 0.017352331429719925, + 0.055936701595783234, + 0.022936558350920677, + 0.017409181222319603, + 0.011718056164681911, + 0.008496779017150402, + 0.03228842467069626, + 0.007848449982702732, + 0.011059204116463661, + 0.035622917115688324, + 0.012623143382370472, + 0.010355993174016476, + 0.007234789896756411, + 0.0008844375261105597, + 0.0035869558341801167, + 0.0007759698783047497, + 0.0013819326413795352, + 0.0036334155593067408, + 0.0015058956341817975, + 0.0012414189986884594, + 0.0009612536523491144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.006739859934896231 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static, slightly high-angle close-up of a cylindrical glass filled with transparent water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, minimalist backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a solid flake of white soapstone at the bottom of the glass begin to dissolve violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do giant silver bubbles erupt from the water surface to form a shimmering vertical stream that rises and coalesces into a dense, pure white foam layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by crisp, clean lighting that emphasizes the clarity of the glass and the texture of the foam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011845960980281234 + ], + "opposite_logprob_tokens": [ + -6.751184463500977 + ], + "logprob_reward": -0.0011845960980281234, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011845960980281234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10558509081602097, + 0.1678113341331482, + 0.06788284331560135, + 0.2622673511505127, + 0.060285720974206924, + 0.07514999806880951, + 0.041960641741752625, + 0.13700951635837555, + 0.0392182320356369, + 0.0428292378783226 + ], + "credit_map": [ + 0.021993402391672134, + 0.010939136147499084, + 0.0077326237224042416, + 0.016278225928544998, + 0.00865761749446392, + 0.008767894469201565, + 0.01742677576839924, + 0.013789418153464794, + 0.011570531874895096, + 0.009224273264408112, + 0.05674002692103386, + 0.03688439726829529, + 0.0066725630313158035, + 0.005150997079908848, + 0.014268501661717892, + 0.02730005607008934, + 0.006814301945269108, + 0.00633603148162365, + 0.006700471974909306, + 0.007972361519932747, + 0.005776623729616404, + 0.00705573707818985, + 0.014741181395947933, + 0.012486133724451065, + 0.03531124070286751, + 0.027184737846255302, + 0.025817405432462692, + 0.03241276368498802, + 0.04629278928041458, + 0.03075050376355648, + 0.0390978641808033, + 0.02540004998445511, + 0.0059698945842683315, + 0.00877736322581768, + 0.005614139139652252, + 0.008851158432662487, + 0.01001647301018238, + 0.005408436991274357, + 0.009306073188781738, + 0.00634218193590641, + 0.006142330355942249, + 0.0074929604306817055, + 0.0068190270103514194, + 0.015057628974318504, + 0.012812149710953236, + 0.007072262465953827, + 0.011662050150334835, + 0.008091593161225319, + 0.0031921714544296265, + 0.004659111145883799, + 0.0045312196016311646, + 0.00593313155695796, + 0.006900880951434374, + 0.0057879225350916386, + 0.005587994121015072, + 0.005368213169276714, + 0.009079224430024624, + 0.02165939100086689, + 0.012412034906446934, + 0.013468913733959198, + 0.039538443088531494, + 0.01775088906288147, + 0.014043520204722881, + 0.009057098999619484, + 0.0025802445597946644, + 0.005727754440158606, + 0.00389497890137136, + 0.004544307943433523, + 0.00799774844199419, + 0.0057397643104195595, + 0.004118633922189474, + 0.00461480300873518, + 0.003948514815419912, + 0.0061536733992397785, + 0.00427038362249732, + 0.0063055832870304585, + 0.00784063246101141, + 0.0038834325969219208, + 0.00486816605553031, + 0.005558847449719906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.0011845960980281234 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_5/index_tensor([2838, 2838])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_5/index_tensor([2838, 2838])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..002e68f63de2fcba8c9e7f4c4500c1fcf7507fc7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_5/index_tensor([2838, 2838])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot shot from an extremely low angle, the camera tilts slowly and steadily upward, following a massive spacecraft, its metal hull covered in weathered titanium panels and a complex network of pipes. Against the deep blue of the predawn sky, it rises with majestic and unstoppable grace from its massive concrete and steel launch pad. At the ship's rear, massive engine nozzles violently belch a dazzling burst of flames. The core of the flames is a near-incandescent white, fading into blazing orange and yellow. Their immense power distorts the surrounding air. Simultaneously, a torrent of thick white smoke and steam erupts from the base, rapidly engulfing the entire launch pad and billowing across the ground, illuminated by the intense light of the flames like a boiling sea of clouds. As the spacecraft steadily and calmly accelerates upward, the camera continues to look up, capturing every structural detail on its huge body illuminated by the fire below. The entire picture is filled with an overwhelming sense of power and an awe-inspiring atmosphere, creating an epic moment of human technology and space exploration.", + "seed": 9609, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_5/index_2838_seed_9609_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06784884221851825, + "vlm_logprob": -0.06784884221851825, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.06784884221851825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2665681540966034 + ], + "opposite_logprob_tokens": [ + -1.5165681838989258 + ], + "logprob_reward": -0.2665681540966034, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2665681540966034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15760967135429382, + 0.04117017239332199, + 0.13873252272605896, + 0.03570520877838135, + 0.10168059170246124, + 0.16050921380519867, + 0.0993918776512146, + 0.08734534680843353, + 0.11466987431049347, + 0.06318552792072296 + ], + "credit_map": [ + 0.01840125396847725, + 0.015593948774039745, + 0.028051361441612244, + 0.031061705201864243, + 0.01377897709608078, + 0.01686927117407322, + 0.014642097987234592, + 0.019211044535040855, + 0.002884653862565756, + 0.0033593312837183475, + 0.00841520819813013, + 0.004880618769675493, + 0.003764787223190069, + 0.007400043308734894, + 0.004780665505677462, + 0.005684861447662115, + 0.00888537522405386, + 0.0061737908981740475, + 0.04463312029838562, + 0.009644179604947567, + 0.004962739069014788, + 0.043197277933359146, + 0.004546415060758591, + 0.01668962463736534, + 0.0042348294518888, + 0.004006260074675083, + 0.004408771172165871, + 0.0033865247387439013, + 0.003838141215965152, + 0.004195487126708031, + 0.005794566124677658, + 0.005840630736202002, + 0.01044618058949709, + 0.008249377831816673, + 0.008174722082912922, + 0.011668077670037746, + 0.006074711214751005, + 0.019856266677379608, + 0.004865170922130346, + 0.03234608843922615, + 0.020766032859683037, + 0.026828082278370857, + 0.014652633108198643, + 0.018543846905231476, + 0.019197281450033188, + 0.031712185591459274, + 0.012215875089168549, + 0.01659327931702137, + 0.009021992795169353, + 0.009279926307499409, + 0.012229073792696, + 0.022545401006937027, + 0.012143377214670181, + 0.009593297727406025, + 0.014946350827813148, + 0.009632463566958904, + 0.023372016847133636, + 0.009467309340834618, + 0.013297529891133308, + 0.0051976582035422325, + 0.007963030599057674, + 0.010760367847979069, + 0.005678033921867609, + 0.011609393171966076, + 0.015587704256176949, + 0.011817330494523048, + 0.018359720706939697, + 0.010809224098920822, + 0.010633234865963459, + 0.02050854079425335, + 0.010807536542415619, + 0.016146577894687653, + 0.00852771382778883, + 0.004363243468105793, + 0.010998599231243134, + 0.006580538582056761, + 0.005687750410288572, + 0.006632540374994278, + 0.006079750135540962, + 0.014315392822027206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.2665681540966034 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036556576378643513 + ], + "opposite_logprob_tokens": [ + -5.628655433654785 + ], + "logprob_reward": -0.0036556576378643513, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036556576378643513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14886806905269623, + 0.04481137543916702, + 0.24001182615756989, + 0.03863992914557457, + 0.14070475101470947, + 0.07497412711381912, + 0.10940820723772049, + 0.057455360889434814, + 0.09184436500072479, + 0.053281962871551514 + ], + "credit_map": [ + 0.021092072129249573, + 0.013132750988006592, + 0.02261362597346306, + 0.03152352198958397, + 0.01262366957962513, + 0.016949333250522614, + 0.013199965469539165, + 0.017733134329319, + 0.004421713761985302, + 0.003780253464356065, + 0.012034162878990173, + 0.007417967077344656, + 0.003297076327726245, + 0.005917769856750965, + 0.0036795560736209154, + 0.004262878093868494, + 0.01488152053207159, + 0.009592326357960701, + 0.08356156200170517, + 0.01547907292842865, + 0.007573920302093029, + 0.07111379504203796, + 0.006987487431615591, + 0.030822133645415306, + 0.004205253906548023, + 0.004706149455159903, + 0.003936081659048796, + 0.0033201887272298336, + 0.005102160386741161, + 0.0051224795170128345, + 0.004805734381079674, + 0.007441881578415632, + 0.0119386687874794, + 0.010635613463819027, + 0.01299211010336876, + 0.018921678885817528, + 0.009504726156592369, + 0.023447787389159203, + 0.015216001309454441, + 0.0380481593310833, + 0.009402722120285034, + 0.01076455693691969, + 0.009215932339429855, + 0.008315720595419407, + 0.009076294489204884, + 0.013410637155175209, + 0.006193808279931545, + 0.008594454266130924, + 0.009586963802576065, + 0.012656143866479397, + 0.013126671314239502, + 0.013986567035317421, + 0.014534985646605492, + 0.011971261352300644, + 0.02209695801138878, + 0.011448643170297146, + 0.010216444730758667, + 0.00591864762827754, + 0.009107546880841255, + 0.004949782509356737, + 0.005304498597979546, + 0.009661905467510223, + 0.0054957992397248745, + 0.006800739094614983, + 0.01351819559931755, + 0.00898183323442936, + 0.011042100377380848, + 0.010196388699114323, + 0.010595380328595638, + 0.015806281939148903, + 0.00986452680081129, + 0.011839663609862328, + 0.007235336117446423, + 0.005964023992419243, + 0.007297436706721783, + 0.00551825063303113, + 0.005382992792874575, + 0.005607143044471741, + 0.005620867013931274, + 0.01065591350197792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0036556576378643513 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635337766259909 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.007635337766259909, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635337766259909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37019750475883484, + 0.02787262760102749, + 0.09746436774730682, + 0.034111328423023224, + 0.16341765224933624, + 0.17630000412464142, + 0.04265983775258064, + 0.029887355864048004, + 0.03222925215959549, + 0.025860073044896126 + ], + "credit_map": [ + 0.032076869159936905, + 0.03632918745279312, + 0.07020191103219986, + 0.10494054108858109, + 0.02695915475487709, + 0.02565300464630127, + 0.03864362835884094, + 0.035393185913562775, + 0.00296348100528121, + 0.0026857347693294287, + 0.006307598669081926, + 0.004596390761435032, + 0.0022523256484419107, + 0.0036501141730695963, + 0.002170260762795806, + 0.00324672251008451, + 0.0060929832980036736, + 0.004048703704029322, + 0.03510912135243416, + 0.00661835353821516, + 0.003469182178378105, + 0.026738930493593216, + 0.003776164259761572, + 0.011610940098762512, + 0.004897364880889654, + 0.005511976312845945, + 0.003792084287852049, + 0.00260849273763597, + 0.0038563290145248175, + 0.004845853429287672, + 0.0036901829298585653, + 0.004909047391265631, + 0.010995232500135899, + 0.012884611263871193, + 0.01102630328387022, + 0.03356350585818291, + 0.007016615476459265, + 0.021731242537498474, + 0.012142007239162922, + 0.05405813455581665, + 0.013947931118309498, + 0.017569774761795998, + 0.021739022806286812, + 0.04029173031449318, + 0.01528715156018734, + 0.015023947693407536, + 0.03095714934170246, + 0.021483320742845535, + 0.003818356664851308, + 0.004247208125889301, + 0.00554294977337122, + 0.0074449572712183, + 0.004934175405651331, + 0.0035541951656341553, + 0.008549136109650135, + 0.004568862728774548, + 0.0056052617728710175, + 0.003539960365742445, + 0.0035450176801532507, + 0.0026412673760205507, + 0.0032866497058421373, + 0.00402632262557745, + 0.0038988052401691675, + 0.0033440738916397095, + 0.004239199683070183, + 0.0031675794161856174, + 0.005221010185778141, + 0.0027986180502921343, + 0.0030833815690129995, + 0.005697137210518122, + 0.0035581267438828945, + 0.004464201163500547, + 0.003998987842351198, + 0.0035213062074035406, + 0.004100808408111334, + 0.0023872319143265486, + 0.0020792223513126373, + 0.002344974083825946, + 0.003216391894966364, + 0.004211151506751776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.007635337766259909 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810269385576248 + ], + "opposite_logprob_tokens": [ + -3.288102626800537 + ], + "logprob_reward": -0.03810269385576248, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810269385576248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06815903633832932, + 0.13540077209472656, + 0.18005722761154175, + 0.039243221282958984, + 0.05068492889404297, + 0.17546707391738892, + 0.1298094391822815, + 0.03700970858335495, + 0.07999489456415176, + 0.10417380183935165 + ], + "credit_map": [ + 0.0058722784742712975, + 0.009967832826077938, + 0.008165520615875721, + 0.012583935633301735, + 0.007755909115076065, + 0.007588494569063187, + 0.00791594386100769, + 0.008309120312333107, + 0.014706647023558617, + 0.013118172064423561, + 0.03462617099285126, + 0.02080605924129486, + 0.00871968362480402, + 0.020527252927422523, + 0.011446595191955566, + 0.011450191959738731, + 0.011574994772672653, + 0.0074735842645168304, + 0.06165974587202072, + 0.011450798250734806, + 0.006411747075617313, + 0.05725729092955589, + 0.00466563506051898, + 0.019563430920243263, + 0.004211479797959328, + 0.005171781871467829, + 0.0039246464148163795, + 0.003149921540170908, + 0.004225492011755705, + 0.004904589150100946, + 0.005089716054499149, + 0.008565593510866165, + 0.00558143574744463, + 0.00532525172457099, + 0.004823663737624884, + 0.009443624876439571, + 0.004903216380625963, + 0.006476497743278742, + 0.005214325152337551, + 0.008916915394365788, + 0.016898950561881065, + 0.017097769305109978, + 0.025092143565416336, + 0.03380831331014633, + 0.016141800209879875, + 0.017818916589021683, + 0.024661831557750702, + 0.02394734136760235, + 0.010559117421507835, + 0.015383997932076454, + 0.015434023924171925, + 0.018955305218696594, + 0.01724558137357235, + 0.012497116811573505, + 0.025102030485868454, + 0.014632259495556355, + 0.006255602929741144, + 0.004502502270042896, + 0.0048691327683627605, + 0.00286848028190434, + 0.004375580232590437, + 0.004536474123597145, + 0.004418082069605589, + 0.0051838550716638565, + 0.012121560052037239, + 0.006958583369851112, + 0.011777318082749844, + 0.008940331637859344, + 0.00960291177034378, + 0.012607738375663757, + 0.006714772433042526, + 0.011271676048636436, + 0.017307253554463387, + 0.009590862318873405, + 0.015616525895893574, + 0.01177817489951849, + 0.011105343699455261, + 0.011577497236430645, + 0.011242464184761047, + 0.015955673530697823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.03810269385576248 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328236773610115 + ], + "opposite_logprob_tokens": [ + -3.773282289505005 + ], + "logprob_reward": -0.02328236773610115, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328236773610115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08425770699977875, + 0.049069587141275406, + 0.2794784605503082, + 0.060964856296777725, + 0.12211225181818008, + 0.08325226604938507, + 0.10932885855436325, + 0.14715683460235596, + 0.026224419474601746, + 0.03815481811761856 + ], + "credit_map": [ + 0.007472085300832987, + 0.009336416609585285, + 0.01124114915728569, + 0.015928372740745544, + 0.007513113785535097, + 0.009286971762776375, + 0.012821540236473083, + 0.010658053681254387, + 0.004024577792733908, + 0.004568369593471289, + 0.009435742162168026, + 0.009037038311362267, + 0.004079393111169338, + 0.009436444379389286, + 0.004214075393974781, + 0.0042739445343613625, + 0.018046267330646515, + 0.012037444859743118, + 0.09445799887180328, + 0.01851656660437584, + 0.009331749752163887, + 0.08673334121704102, + 0.007828954607248306, + 0.03252609819173813, + 0.007598217576742172, + 0.011460414156317711, + 0.006592675112187862, + 0.005179706495255232, + 0.00783915538340807, + 0.006847687065601349, + 0.00677528278902173, + 0.0086717139929533, + 0.007586363703012466, + 0.008853055536746979, + 0.013345319777727127, + 0.027208128944039345, + 0.008016908541321754, + 0.012228445149958134, + 0.01077500730752945, + 0.034099020063877106, + 0.012054174207150936, + 0.012147346511483192, + 0.010145405307412148, + 0.01019988115876913, + 0.010286766104400158, + 0.011276978999376297, + 0.007639806251972914, + 0.009501904249191284, + 0.013289892114698887, + 0.01480770017951727, + 0.0171864815056324, + 0.013842934742569923, + 0.011217976920306683, + 0.01824132725596428, + 0.009695055894553661, + 0.011047475039958954, + 0.022740477696061134, + 0.02251429110765457, + 0.02869623340666294, + 0.009378467686474323, + 0.014848877675831318, + 0.01624476909637451, + 0.00824340432882309, + 0.024490300565958023, + 0.0033332540187984705, + 0.002979044569656253, + 0.0034022051841020584, + 0.0033283925149589777, + 0.004014451056718826, + 0.0032400512136518955, + 0.0026432049926370382, + 0.0032838142942637205, + 0.005597976967692375, + 0.004103589802980423, + 0.004672477021813393, + 0.004254297353327274, + 0.003751991782337427, + 0.004283852409571409, + 0.0044434210285544395, + 0.007047206629067659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.02328236773610115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera tilts slowly and steadily upward, following a massive spacecraft, its metal hull covered in weathered titanium panels and a complex network of pipes. Against the deep blue of the predawn sky, it rises with majestic and unstoppable grace from its massive concrete and steel launch pad. At the ship's rear, massive engine nozzles violently belch a dazzling burst of flames. The core of the flames is a near-incandescent white, fading into blazing orange and yellow. Their immense power distorts the surrounding air. Simultaneously, a torrent of thick white smoke and steam erupts from the base, rapidly engulfing the entire launch pad and billowing across the ground, illuminated by the intense light of the flames like a boiling sea of clouds. As the spacecraft steadily and calmly accelerates upward, the camera continues to look up, capturing every structural detail on its huge body illuminated by the fire below. The entire picture is filled with an overwhelming sense of power and an awe-inspiring atmosphere, creating an epic moment of human technology and space exploration.", + "seed": 9610, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_5/index_2838_seed_9610_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03315356438979507, + "vlm_logprob": -0.03315356438979507, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.03315356438979507, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1098305955529213 + ], + "opposite_logprob_tokens": [ + -2.359830617904663 + ], + "logprob_reward": -0.1098305955529213, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1098305955529213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19306647777557373, + 0.2243015170097351, + 0.048021361231803894, + 0.053912412375211716, + 0.0763663575053215, + 0.08133168518543243, + 0.04521406814455986, + 0.07504122704267502, + 0.05407882481813431, + 0.14866605401039124 + ], + "credit_map": [ + 0.024382922798395157, + 0.02328009530901909, + 0.026962490752339363, + 0.023574702441692352, + 0.019674938172101974, + 0.019284294918179512, + 0.027123888954520226, + 0.028783133253455162, + 0.008896795101463795, + 0.01378966961055994, + 0.01721845008432865, + 0.041010499000549316, + 0.016875391826033592, + 0.01720481552183628, + 0.05970229580998421, + 0.04960360750555992, + 0.0023103419225662947, + 0.004762775730341673, + 0.003578920615836978, + 0.014293167740106583, + 0.00539338868111372, + 0.003493997035548091, + 0.006372026167809963, + 0.007816746830940247, + 0.005871918518096209, + 0.006724428851157427, + 0.005009831860661507, + 0.005819510668516159, + 0.010976981371641159, + 0.005241020116955042, + 0.007259808015078306, + 0.007008910179138184, + 0.005240061320364475, + 0.009778306819498539, + 0.0057130069471895695, + 0.004762259777635336, + 0.023802028968930244, + 0.009024386294186115, + 0.007380760740488768, + 0.010665543377399445, + 0.010601837188005447, + 0.012541037052869797, + 0.006128306966274977, + 0.008017629384994507, + 0.014027206227183342, + 0.008543141186237335, + 0.010990966111421585, + 0.010481559671461582, + 0.005866596475243568, + 0.0049033029936254025, + 0.006079173181205988, + 0.0056658657267689705, + 0.0032950397580862045, + 0.005840470548719168, + 0.006383139174431562, + 0.007180481217801571, + 0.0058027287013828754, + 0.01633143238723278, + 0.005224703811109066, + 0.005103344097733498, + 0.010110177099704742, + 0.012251575477421284, + 0.007393689826130867, + 0.012823574244976044, + 0.005940987728536129, + 0.007804775610566139, + 0.005465611349791288, + 0.007790166884660721, + 0.005332210101187229, + 0.008050989359617233, + 0.007740064524114132, + 0.005954019725322723, + 0.010179487988352776, + 0.0053052450530231, + 0.0055047934874892235, + 0.004093718249350786, + 0.006638843100517988, + 0.05491406470537186, + 0.0047571733593940735, + 0.05727270990610123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.1098305955529213 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004121382255107164 + ], + "opposite_logprob_tokens": [ + -5.50412130355835 + ], + "logprob_reward": -0.004121382255107164, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004121382255107164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08860568702220917, + 0.11803998053073883, + 0.09320011734962463, + 0.08507731556892395, + 0.07909185439348221, + 0.043449707329273224, + 0.04211388900876045, + 0.04953283816576004, + 0.03144003078341484, + 0.36944860219955444 + ], + "credit_map": [ + 0.009177607484161854, + 0.011045219376683235, + 0.009935314767062664, + 0.015744129195809364, + 0.00852839183062315, + 0.007116775959730148, + 0.012076979503035545, + 0.014981270767748356, + 0.006015917286276817, + 0.009094211272895336, + 0.00704685365781188, + 0.022288838401436806, + 0.009994518011808395, + 0.00941616203635931, + 0.030798211693763733, + 0.023385275155305862, + 0.004721149802207947, + 0.008473845198750496, + 0.006333981174975634, + 0.015798384323716164, + 0.014702790416777134, + 0.01136116124689579, + 0.01817985624074936, + 0.013628941029310226, + 0.010369031690061092, + 0.0084050502628088, + 0.009187639690935612, + 0.010075820609927177, + 0.011192506179213524, + 0.007373754866421223, + 0.013220040127635002, + 0.01525347214192152, + 0.004487983416765928, + 0.008110839873552322, + 0.005828403867781162, + 0.003195220837369561, + 0.026506368070840836, + 0.006861439906060696, + 0.005766244139522314, + 0.01833534613251686, + 0.005615968257188797, + 0.006624642293900251, + 0.0035907786805182695, + 0.004523003939539194, + 0.006528418976813555, + 0.005650929640978575, + 0.004470657557249069, + 0.006445304024964571, + 0.0061049447394907475, + 0.005729925353080034, + 0.005377288907766342, + 0.006091875955462456, + 0.0035144216381013393, + 0.004660617560148239, + 0.004598397295922041, + 0.006036415696144104, + 0.0039601256139576435, + 0.011004878208041191, + 0.004114129114896059, + 0.0030323106329888105, + 0.004404699895530939, + 0.00781954638659954, + 0.004292125813663006, + 0.010905024595558643, + 0.003578523639589548, + 0.005832568742334843, + 0.0031941384077072144, + 0.002659644465893507, + 0.0032646339386701584, + 0.00457161758095026, + 0.004162754397839308, + 0.004176150541752577, + 0.027316946536302567, + 0.009560519829392433, + 0.010685835033655167, + 0.008623958565294743, + 0.013866632245481014, + 0.14208117127418518, + 0.008587050251662731, + 0.14872652292251587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.004121382255107164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951658822596073 + ], + "opposite_logprob_tokens": [ + -5.130951881408691 + ], + "logprob_reward": -0.005951658822596073, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951658822596073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18713533878326416, + 0.14180247485637665, + 0.09161104261875153, + 0.05714551731944084, + 0.09631796181201935, + 0.03962015360593796, + 0.02559180185198784, + 0.02732623741030693, + 0.036074500530958176, + 0.2973749041557312 + ], + "credit_map": [ + 0.01630973443388939, + 0.021552961319684982, + 0.020917536690831184, + 0.0313158854842186, + 0.01340755820274353, + 0.0155125567689538, + 0.03663454204797745, + 0.031484585255384445, + 0.005279232282191515, + 0.008117583580315113, + 0.01117882039397955, + 0.026580462232232094, + 0.00966047402471304, + 0.008254793472588062, + 0.04017806425690651, + 0.0325530469417572, + 0.003811474423855543, + 0.010013071820139885, + 0.005457446910440922, + 0.01822087913751602, + 0.01330835372209549, + 0.0115323755890131, + 0.014380642212927341, + 0.014886811375617981, + 0.006116385571658611, + 0.006269157864153385, + 0.00497076939791441, + 0.007656640838831663, + 0.007183459587395191, + 0.006028477102518082, + 0.009045695886015892, + 0.009874935261905193, + 0.005545940715819597, + 0.010268892161548138, + 0.0073733036406338215, + 0.003916923422366381, + 0.031793732196092606, + 0.008069552481174469, + 0.007824478670954704, + 0.021525152027606964, + 0.006860330235213041, + 0.0053289588540792465, + 0.003404054557904601, + 0.004223507829010487, + 0.004841099958866835, + 0.004486272111535072, + 0.005067422520369291, + 0.005408511031419039, + 0.002081942278891802, + 0.0027768879663199186, + 0.002325876150280237, + 0.0031856412533670664, + 0.003847871208563447, + 0.0036018460523337126, + 0.0029180787969380617, + 0.004853660240769386, + 0.0030927981715649366, + 0.00695362314581871, + 0.0022940970957279205, + 0.001569531043060124, + 0.0027319409418851137, + 0.00423054676502943, + 0.001745139597915113, + 0.004708561580628157, + 0.004050809424370527, + 0.005268694832921028, + 0.004117507487535477, + 0.003791090566664934, + 0.0027753161266446114, + 0.005520780570805073, + 0.004843881353735924, + 0.005706424359232187, + 0.0239909328520298, + 0.006859842222183943, + 0.0079533401876688, + 0.0059832320548594, + 0.009260670281946659, + 0.11557618528604507, + 0.006540222093462944, + 0.12121051549911499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.005951658822596073 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071418300271034 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071418300271034, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071418300271034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07240136712789536, + 0.07344681769609451, + 0.033606160432100296, + 0.05161906033754349, + 0.16604264080524445, + 0.026104707270860672, + 0.016504796221852303, + 0.036995235830545425, + 0.05422114580869675, + 0.4690580368041992 + ], + "credit_map": [ + 0.008540034294128418, + 0.010307866148650646, + 0.009166033007204533, + 0.008088130503892899, + 0.007000994868576527, + 0.007419419940561056, + 0.012136707082390785, + 0.00974217988550663, + 0.0034504563082009554, + 0.0051712533459067345, + 0.0057248277589678764, + 0.013027306646108627, + 0.006854122970253229, + 0.006898061838001013, + 0.017960766330361366, + 0.01436002179980278, + 0.0014766729436814785, + 0.0039061512798070908, + 0.0021190939005464315, + 0.006485200021415949, + 0.0038656212855130434, + 0.00536147877573967, + 0.005172585137188435, + 0.005219355691224337, + 0.006581260822713375, + 0.007550379261374474, + 0.007717763539403677, + 0.0053044590167701244, + 0.008897736668586731, + 0.004567589145153761, + 0.005941987503319979, + 0.005057887639850378, + 0.01893693208694458, + 0.015144987031817436, + 0.03627493977546692, + 0.01044414471834898, + 0.039559219032526016, + 0.013109564781188965, + 0.014559953473508358, + 0.01801290363073349, + 0.0026553424540907145, + 0.004099678713828325, + 0.0018806875450536609, + 0.0026716638822108507, + 0.004219988826662302, + 0.0033075259998440742, + 0.004005548544228077, + 0.003264271654188633, + 0.0019280794076621532, + 0.0015618245815858245, + 0.001217362587340176, + 0.0019346941262483597, + 0.0015558248851448298, + 0.0023503273259848356, + 0.0019410080276429653, + 0.004015675280243158, + 0.0034783666487783194, + 0.005773505661636591, + 0.003307902719825506, + 0.002361912280321121, + 0.004777328576892614, + 0.0064702387899160385, + 0.002831193618476391, + 0.007994786836206913, + 0.006537914276123047, + 0.008276932872831821, + 0.0051292539574205875, + 0.005010965745896101, + 0.0037655653432011604, + 0.009048660285770893, + 0.005688661243766546, + 0.010763194411993027, + 0.0411723293364048, + 0.015090443193912506, + 0.015216483734548092, + 0.013661359436810017, + 0.018871650099754333, + 0.1714174449443817, + 0.013350157998502254, + 0.18027819693088531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.016071418300271034 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792767018079758 + ], + "opposite_logprob_tokens": [ + -3.5297927856445312 + ], + "logprob_reward": -0.029792767018079758, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792767018079758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1423170417547226, + 0.11218121647834778, + 0.0716143324971199, + 0.08327272534370422, + 0.04956085979938507, + 0.08044559508562088, + 0.06967251002788544, + 0.07696318626403809, + 0.06726489216089249, + 0.24670763313770294 + ], + "credit_map": [ + 0.013108402490615845, + 0.01022266037762165, + 0.0123361861333251, + 0.014575377106666565, + 0.010262464173138142, + 0.007712001446634531, + 0.04740682989358902, + 0.026693133637309074, + 0.006813778541982174, + 0.009353119879961014, + 0.009005825035274029, + 0.01982771046459675, + 0.010886279866099358, + 0.013107719831168652, + 0.02424580231308937, + 0.018940988928079605, + 0.003399825422093272, + 0.007789938244968653, + 0.004194074776023626, + 0.01626536063849926, + 0.00894141849130392, + 0.008853544481098652, + 0.01041534822434187, + 0.011754830367863178, + 0.013599895872175694, + 0.010324993170797825, + 0.012419324368238449, + 0.007573367096483707, + 0.01006261631846428, + 0.00753813749179244, + 0.00986038614064455, + 0.011894017457962036, + 0.005456054117530584, + 0.00520846014842391, + 0.008282472379505634, + 0.005340777803212404, + 0.0069258674047887325, + 0.0038167773745954037, + 0.005525969434529543, + 0.009004488587379456, + 0.007092669606208801, + 0.013940079137682915, + 0.007032064720988274, + 0.00989566370844841, + 0.012524626217782497, + 0.009714349173009396, + 0.010293034836649895, + 0.00995311513543129, + 0.011353613808751106, + 0.00911946315318346, + 0.006115144118666649, + 0.009963206015527248, + 0.007096551824361086, + 0.007238311693072319, + 0.009374363347887993, + 0.009411866776645184, + 0.007132051046937704, + 0.019119493663311005, + 0.007277748081833124, + 0.004835599567741156, + 0.011367611587047577, + 0.010419860482215881, + 0.00784342736005783, + 0.008967394009232521, + 0.008152064867317677, + 0.007923985831439495, + 0.008929342031478882, + 0.008985113352537155, + 0.008153954520821571, + 0.00855175033211708, + 0.007961496710777283, + 0.008607183583080769, + 0.024485832080245018, + 0.011625966057181358, + 0.013788319192826748, + 0.009093083441257477, + 0.014498758129775524, + 0.08008173853158951, + 0.009125426411628723, + 0.08400852978229523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.029792767018079758 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera tilts slowly and steadily upward, following a massive spacecraft, its metal hull covered in weathered titanium panels and a complex network of pipes. Against the deep blue of the predawn sky, it rises with majestic and unstoppable grace from its massive concrete and steel launch pad. At the ship's rear, massive engine nozzles violently belch a dazzling burst of flames. The core of the flames is a near-incandescent white, fading into blazing orange and yellow. Their immense power distorts the surrounding air. Simultaneously, a torrent of thick white smoke and steam erupts from the base, rapidly engulfing the entire launch pad and billowing across the ground, illuminated by the intense light of the flames like a boiling sea of clouds. As the spacecraft steadily and calmly accelerates upward, the camera continues to look up, capturing every structural detail on its huge body illuminated by the fire below. The entire picture is filled with an overwhelming sense of power and an awe-inspiring atmosphere, creating an epic moment of human technology and space exploration.", + "seed": 9611, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_5/index_2838_seed_9611_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16149236322380603, + "vlm_logprob": -0.16149236322380603, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.16149236322380603, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.717476487159729 + ], + "opposite_logprob_tokens": [ + -0.717476487159729 + ], + "logprob_reward": -0.717476487159729, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.717476487159729, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13904330134391785, + 0.05476038530468941, + 0.027120862156152725, + 0.10535044968128204, + 0.044292498379945755, + 0.04680072143673897, + 0.3230876326560974, + 0.0917583480477333, + 0.11660940200090408, + 0.051176466047763824 + ], + "credit_map": [ + 0.020841404795646667, + 0.014076617546379566, + 0.020766936242580414, + 0.017333734780550003, + 0.01772216521203518, + 0.01834212802350521, + 0.016357671469449997, + 0.01360264327377081, + 0.007148980163037777, + 0.005292392335832119, + 0.007526970002800226, + 0.004409893415868282, + 0.0062349834479391575, + 0.012860088609158993, + 0.0048334295861423016, + 0.0064536454156041145, + 0.005449203308671713, + 0.003802814055234194, + 0.0028735597152262926, + 0.0024684234522283077, + 0.0032278341241180897, + 0.004580838140100241, + 0.00202357629314065, + 0.002694613067433238, + 0.00960185844451189, + 0.008838805370032787, + 0.0034580850042402744, + 0.00459691509604454, + 0.0034735840745270252, + 0.007812196854501963, + 0.026493703946471214, + 0.04107530042529106, + 0.005010557826608419, + 0.00404277304187417, + 0.0037654631305485964, + 0.004021139349788427, + 0.002954841824248433, + 0.003085839096456766, + 0.008898776955902576, + 0.012513107620179653, + 0.006037871818989515, + 0.006000224966555834, + 0.004116509109735489, + 0.006657228339463472, + 0.005262267775833607, + 0.00424485607072711, + 0.007423363626003265, + 0.007058399263769388, + 0.02415398322045803, + 0.026490909978747368, + 0.014264577068388462, + 0.08536836504936218, + 0.02989860810339451, + 0.01423267275094986, + 0.09926381707191467, + 0.0294147077947855, + 0.010062719695270061, + 0.012776163406670094, + 0.010016858577728271, + 0.010745485313236713, + 0.019647788256406784, + 0.009573844261467457, + 0.010795150883495808, + 0.00814033206552267, + 0.009160489775240421, + 0.01662750542163849, + 0.009052673354744911, + 0.01608683541417122, + 0.020798334851861, + 0.011216205544769764, + 0.011984808370471, + 0.021682551130652428, + 0.00454295426607132, + 0.00814005732536316, + 0.007072812411934137, + 0.005335914436727762, + 0.007553865201771259, + 0.004433084279298782, + 0.008355310186743736, + 0.00574246933683753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.717476487159729 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028746493626385927 + ], + "opposite_logprob_tokens": [ + -5.877874851226807 + ], + "logprob_reward": -0.0028746493626385927, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028746493626385927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13959001004695892, + 0.03568044304847717, + 0.046969518065452576, + 0.03267138451337814, + 0.047189243137836456, + 0.040195807814598083, + 0.07150334119796753, + 0.18265755474567413, + 0.292611300945282, + 0.11093137413263321 + ], + "credit_map": [ + 0.017787892371416092, + 0.015634829178452492, + 0.0208198893815279, + 0.018689442425966263, + 0.016123365610837936, + 0.01809067651629448, + 0.019069993868470192, + 0.013373913243412971, + 0.004486425314098597, + 0.004459771327674389, + 0.005669759586453438, + 0.0047263759188354015, + 0.003927293233573437, + 0.006022584158927202, + 0.0029388878028839827, + 0.0034493443090468645, + 0.006548077799379826, + 0.005486563313752413, + 0.004912750795483589, + 0.007678360212594271, + 0.004822961054742336, + 0.00603362824767828, + 0.003583208192139864, + 0.00790396612137556, + 0.00544591573998332, + 0.002476704539731145, + 0.0022267226595431566, + 0.0022688687313348055, + 0.0016371938399970531, + 0.002962157828733325, + 0.005505874287337065, + 0.010147946886718273, + 0.0052553704008460045, + 0.005455965641885996, + 0.005298807751387358, + 0.005231140181422234, + 0.004023318644613028, + 0.0036750624421983957, + 0.008487549610435963, + 0.00976202730089426, + 0.0046772402711212635, + 0.005579098593443632, + 0.003861626610159874, + 0.004814657848328352, + 0.005527406465262175, + 0.003636718261986971, + 0.0048151216469705105, + 0.007283938117325306, + 0.005000209901481867, + 0.004681678023189306, + 0.0029105618596076965, + 0.021579723805189133, + 0.00582914799451828, + 0.0028931733686476946, + 0.022378072142601013, + 0.006230780389159918, + 0.01945745199918747, + 0.009067093953490257, + 0.005651002284139395, + 0.01439419761300087, + 0.015373439528048038, + 0.011733573861420155, + 0.054367437958717346, + 0.05261336639523506, + 0.04260440915822983, + 0.019257649779319763, + 0.042295586317777634, + 0.02559257671236992, + 0.05217193439602852, + 0.025270521640777588, + 0.022152671590447426, + 0.06326594948768616, + 0.01131840143352747, + 0.015008502639830112, + 0.011422495357692242, + 0.012344222515821457, + 0.016388434916734695, + 0.011951250955462456, + 0.02068405970931053, + 0.011814008466899395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0028746493626385927 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651277050375938 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651277050375938, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651277050375938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1023227646946907, + 0.0676250234246254, + 0.04746690392494202, + 0.33998948335647583, + 0.05570989474654198, + 0.08216976374387741, + 0.043900053948163986, + 0.08861452341079712, + 0.06913953274488449, + 0.10306208580732346 + ], + "credit_map": [ + 0.01100329402834177, + 0.012208595871925354, + 0.019013555720448494, + 0.015786172822117805, + 0.009296421892940998, + 0.010708203539252281, + 0.012223605066537857, + 0.01208291295915842, + 0.007529071532189846, + 0.00703815184533596, + 0.00824054516851902, + 0.00789701472967863, + 0.011556734330952168, + 0.009270049631595612, + 0.008204474113881588, + 0.007888981141149998, + 0.005758448038250208, + 0.005709065590053797, + 0.004497793968766928, + 0.006450493820011616, + 0.005526326131075621, + 0.006952018942683935, + 0.004098129458725452, + 0.008474621921777725, + 0.029497133567929268, + 0.02645946852862835, + 0.008754701353609562, + 0.012098620645701885, + 0.00946870818734169, + 0.02389993518590927, + 0.09213270992040634, + 0.13767816126346588, + 0.004829219542443752, + 0.006181139964610338, + 0.005025498103350401, + 0.005365852266550064, + 0.004222465213388205, + 0.0042070006020367146, + 0.011629128828644753, + 0.014249592088162899, + 0.008630672469735146, + 0.007664763834327459, + 0.00800779927521944, + 0.013158367946743965, + 0.0084450114518404, + 0.005784568376839161, + 0.012362752109766006, + 0.0181158147752285, + 0.007539251819252968, + 0.005478850100189447, + 0.003912425599992275, + 0.00735529325902462, + 0.005878638941794634, + 0.003141634166240692, + 0.006434949114918709, + 0.004159011412411928, + 0.008842366747558117, + 0.007569728419184685, + 0.004631207324564457, + 0.00742840301245451, + 0.010648166760802269, + 0.005585921928286552, + 0.02320021763443947, + 0.020708506926894188, + 0.007245918735861778, + 0.006544789765030146, + 0.009187736548483372, + 0.008243401534855366, + 0.01180430967360735, + 0.005297978408634663, + 0.008436071686446667, + 0.012379320338368416, + 0.011237844824790955, + 0.013828347437083721, + 0.013909000903367996, + 0.011213935911655426, + 0.01528598740696907, + 0.010912741534411907, + 0.016893822699785233, + 0.00978040136396885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.008651277050375938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866408929228783 + ], + "opposite_logprob_tokens": [ + -3.048664093017578 + ], + "logprob_reward": -0.04866408929228783, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866408929228783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17488500475883484, + 0.24437306821346283, + 0.029479211196303368, + 0.10656555742025375, + 0.045524872839450836, + 0.015581020154058933, + 0.17300210893154144, + 0.10392443835735321, + 0.08311739563941956, + 0.02354729175567627 + ], + "credit_map": [ + 0.027923990041017532, + 0.02663532830774784, + 0.025356251746416092, + 0.019459767267107964, + 0.024216510355472565, + 0.020006800070405006, + 0.012581839226186275, + 0.018704524263739586, + 0.037968892604112625, + 0.026626165956258774, + 0.03530986234545708, + 0.019308211281895638, + 0.03345180302858353, + 0.04759453609585762, + 0.017117060720920563, + 0.02699652872979641, + 0.005272557493299246, + 0.0026832877192646265, + 0.003496823599562049, + 0.004552728962153196, + 0.002644224325194955, + 0.0045823161490261555, + 0.0021933591924607754, + 0.00405391538515687, + 0.008267785422503948, + 0.008678478188812733, + 0.0029061827808618546, + 0.003655536100268364, + 0.0031608243007212877, + 0.008429544046521187, + 0.029201699420809746, + 0.04226551204919815, + 0.004632388241589069, + 0.004396370146423578, + 0.0029899703804403543, + 0.004100719001144171, + 0.002740983385592699, + 0.0028328425250947475, + 0.010167814791202545, + 0.01366378553211689, + 0.002442964818328619, + 0.0023389861453324556, + 0.0015635411255061626, + 0.0017718672752380371, + 0.0021120989695191383, + 0.0017973202047869563, + 0.0015758691588416696, + 0.001978372223675251, + 0.012512576766312122, + 0.014458280056715012, + 0.007395036518573761, + 0.04628570005297661, + 0.016598813235759735, + 0.007593902293592691, + 0.052489280700683594, + 0.015668528154492378, + 0.010742549784481525, + 0.007562211249023676, + 0.00414448045194149, + 0.008075383491814137, + 0.012539003975689411, + 0.007191430777311325, + 0.027385588735342026, + 0.02628379315137863, + 0.008981280960142612, + 0.006935908924788237, + 0.008804794400930405, + 0.008681404404342175, + 0.016237223520874977, + 0.005773844197392464, + 0.0065767802298069, + 0.021126162260770798, + 0.0024737396743148565, + 0.00294463406316936, + 0.002596070524305105, + 0.0032857947517186403, + 0.003845710540190339, + 0.001940616755746305, + 0.0039015975780785084, + 0.0025591268204152584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.04866408929228783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029795313253998756 + ], + "opposite_logprob_tokens": [ + -3.5297954082489014 + ], + "logprob_reward": -0.029795313253998756, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029795313253998756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08119666576385498, + 0.04969318211078644, + 0.04628145694732666, + 0.076873779296875, + 0.06936262547969818, + 0.025956660509109497, + 0.09939951449632645, + 0.09742458164691925, + 0.42637041211128235, + 0.027441129088401794 + ], + "credit_map": [ + 0.009046762250363827, + 0.008275855332612991, + 0.012106405571103096, + 0.011093112640082836, + 0.011020060628652573, + 0.009896285831928253, + 0.01196786854416132, + 0.007790322415530682, + 0.008082191459834576, + 0.004251462873071432, + 0.006250616163015366, + 0.0035059270448982716, + 0.006294147577136755, + 0.0117407301440835, + 0.004363859072327614, + 0.005204247776418924, + 0.008420523256063461, + 0.003214057767763734, + 0.004634270444512367, + 0.006258937995880842, + 0.0041415272280573845, + 0.005847904831171036, + 0.004094784148037434, + 0.009669451974332333, + 0.0050053964368999004, + 0.0058512454852461815, + 0.0017571880016475916, + 0.003187798894941807, + 0.002928987378254533, + 0.006192585453391075, + 0.022367702797055244, + 0.02958287112414837, + 0.006813939660787582, + 0.008677510544657707, + 0.00543485302478075, + 0.007146615535020828, + 0.005753554403781891, + 0.006210807245224714, + 0.01314570289105177, + 0.0161796435713768, + 0.0032238380517810583, + 0.0028318592812865973, + 0.002404264872893691, + 0.004019029438495636, + 0.002825438277795911, + 0.001820332370698452, + 0.004712661728262901, + 0.0041192383505403996, + 0.008320202119648457, + 0.01053752563893795, + 0.0047986675053834915, + 0.02309657819569111, + 0.01205221377313137, + 0.0053201778791844845, + 0.02604493871331215, + 0.009229207411408424, + 0.009723437018692493, + 0.010478931479156017, + 0.006775646470487118, + 0.00909541454166174, + 0.018249912187457085, + 0.007472477853298187, + 0.0185158159583807, + 0.017112940549850464, + 0.04297146573662758, + 0.027593934908509254, + 0.058629363775253296, + 0.036056894809007645, + 0.08944173902273178, + 0.031885042786598206, + 0.030088527128100395, + 0.10970339924097061, + 0.0026729975361377, + 0.003828237298876047, + 0.003067299257963896, + 0.003203820437192917, + 0.0046151005662977695, + 0.002599847735837102, + 0.004469061270356178, + 0.0029847666155546904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.029795313253998756 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from an extremely low angle, the camera tilts slowly and steadily upward, following a massive spacecraft, its metal hull covered in weathered titanium panels and a complex network of pipes. Against the deep blue of the predawn sky, it rises with majestic and unstoppable grace from its massive concrete and steel launch pad. At the ship's rear, massive engine nozzles violently belch a dazzling burst of flames. The core of the flames is a near-incandescent white, fading into blazing orange and yellow. Their immense power distorts the surrounding air. Simultaneously, a torrent of thick white smoke and steam erupts from the base, rapidly engulfing the entire launch pad and billowing across the ground, illuminated by the intense light of the flames like a boiling sea of clouds. As the spacecraft steadily and calmly accelerates upward, the camera continues to look up, capturing every structural detail on its huge body illuminated by the fire below. The entire picture is filled with an overwhelming sense of power and an awe-inspiring atmosphere, creating an epic moment of human technology and space exploration.", + "seed": 9612, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_5/index_2838_seed_9612_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03992476277053356, + "vlm_logprob": -0.03992476277053356, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03992476277053356, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0967148020863533 + ], + "opposite_logprob_tokens": [ + -2.596714735031128 + ], + "logprob_reward": -0.0967148020863533, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0967148020863533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12045642733573914, + 0.06907974928617477, + 0.09868530929088593, + 0.4061741232872009, + 0.051059652119874954, + 0.11564597487449646, + 0.053551044315099716, + 0.01844862848520279, + 0.02711290493607521, + 0.039786115288734436 + ], + "credit_map": [ + 0.010992741212248802, + 0.017659060657024384, + 0.018239639699459076, + 0.016722282394766808, + 0.014954596757888794, + 0.01037298608571291, + 0.016854221001267433, + 0.014660913497209549, + 0.006920722778886557, + 0.009252779185771942, + 0.012341600842773914, + 0.010806571692228317, + 0.007876445539295673, + 0.006450574845075607, + 0.007824412547051907, + 0.007606646046042442, + 0.011296800337731838, + 0.021179217845201492, + 0.00868785660713911, + 0.010787113569676876, + 0.015716612339019775, + 0.012928417883813381, + 0.008320551365613937, + 0.009768751449882984, + 0.0411992222070694, + 0.03942097723484039, + 0.04477037861943245, + 0.09438373893499374, + 0.04328157752752304, + 0.04566989094018936, + 0.04515417292714119, + 0.05229418724775314, + 0.005664489232003689, + 0.006846135016530752, + 0.010012678802013397, + 0.006618179380893707, + 0.005401052534580231, + 0.005763587076216936, + 0.005330061074346304, + 0.0054234727285802364, + 0.00878534372895956, + 0.006890518125146627, + 0.008636700920760632, + 0.026794953271746635, + 0.012658786959946156, + 0.026676366105675697, + 0.010372109711170197, + 0.014831207692623138, + 0.0048746331594884396, + 0.004413130227476358, + 0.0061561670154333115, + 0.009401842951774597, + 0.008036866784095764, + 0.008804797194898129, + 0.00575205497443676, + 0.006111555732786655, + 0.002226033480837941, + 0.0028561048675328493, + 0.0025269503239542246, + 0.0020285749342292547, + 0.0036581309977918863, + 0.0018201213097199798, + 0.002178120194002986, + 0.0011545941233634949, + 0.0031586424447596073, + 0.005574135109782219, + 0.0029209654312580824, + 0.0029553454369306564, + 0.003982136491686106, + 0.002808017423376441, + 0.0031447315122932196, + 0.002568932482972741, + 0.005857240408658981, + 0.006111453287303448, + 0.0058151111006736755, + 0.0036823574919253588, + 0.003720542648807168, + 0.0037309499457478523, + 0.005442134570330381, + 0.005426329094916582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.0967148020863533 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002854917198419571 + ], + "opposite_logprob_tokens": [ + -5.877854824066162 + ], + "logprob_reward": -0.002854917198419571, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002854917198419571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0612223818898201, + 0.043913036584854126, + 0.04536717012524605, + 0.0989864394068718, + 0.13197338581085205, + 0.510364830493927, + 0.056130874902009964, + 0.014223318547010422, + 0.012569244019687176, + 0.02524932473897934 + ], + "credit_map": [ + 0.005179332569241524, + 0.007856539450585842, + 0.006591502111405134, + 0.005344755947589874, + 0.010026631876826286, + 0.006671980954706669, + 0.008349822834134102, + 0.011201811954379082, + 0.004200228024274111, + 0.003904767567291856, + 0.005989063531160355, + 0.004486399237066507, + 0.005896425805985928, + 0.004227628465741873, + 0.006896965205669403, + 0.008311555720865726, + 0.005460487678647041, + 0.006401168182492256, + 0.006486878264695406, + 0.006226758472621441, + 0.0043026902712881565, + 0.005260380916297436, + 0.004143559839576483, + 0.007085245568305254, + 0.01121470332145691, + 0.01171537023037672, + 0.012168139219284058, + 0.013772626407444477, + 0.01582171954214573, + 0.014748857356607914, + 0.008793951012194157, + 0.010751074180006981, + 0.010465930216014385, + 0.01417656522244215, + 0.01569761149585247, + 0.013485279865562916, + 0.014439426362514496, + 0.01747172884643078, + 0.01457459107041359, + 0.03166224807500839, + 0.04040634259581566, + 0.033458296209573746, + 0.04087383672595024, + 0.13635197281837463, + 0.04961176589131355, + 0.13408365845680237, + 0.04655815660953522, + 0.02902083471417427, + 0.005354452412575483, + 0.005013126879930496, + 0.0062504978850483894, + 0.009743801318109035, + 0.00842288974672556, + 0.008983210660517216, + 0.005937767215073109, + 0.006425126921385527, + 0.0016140660736709833, + 0.001637827605009079, + 0.001907023717649281, + 0.001795108080841601, + 0.002311429474502802, + 0.0016095051541924477, + 0.001780161401256919, + 0.0015681972727179527, + 0.0014131504576653242, + 0.002136872848495841, + 0.0011967089958488941, + 0.0016292276559397578, + 0.0017473162151873112, + 0.0017174540553241968, + 0.0014206303749233484, + 0.0013078827178105712, + 0.0035026674158871174, + 0.005288458429276943, + 0.0035902324598282576, + 0.002703983336687088, + 0.002745122881606221, + 0.0024819541722536087, + 0.00287392339669168, + 0.00206298241391778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.002854917198419571 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746135652065277 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746135652065277, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746135652065277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09707964211702347, + 0.2654268443584442, + 0.11437083780765533, + 0.1423216015100479, + 0.09652546048164368, + 0.06907674670219421, + 0.08318301290273666, + 0.025055252015590668, + 0.024780312553048134, + 0.08218027651309967 + ], + "credit_map": [ + 0.00915613304823637, + 0.012508934363722801, + 0.01193577703088522, + 0.01106945239007473, + 0.015200653113424778, + 0.010056511498987675, + 0.013637326657772064, + 0.013514849357306957, + 0.02151045762002468, + 0.030385024845600128, + 0.02526084892451763, + 0.024052659049630165, + 0.03182322904467583, + 0.027969129383563995, + 0.058287933468818665, + 0.046137578785419464, + 0.014272291213274002, + 0.012460191734135151, + 0.015035862103104591, + 0.025939954444766045, + 0.009641404263675213, + 0.01184367761015892, + 0.013935557566583157, + 0.011241895146667957, + 0.01235565822571516, + 0.01309914980083704, + 0.013593446463346481, + 0.02322206273674965, + 0.016393713653087616, + 0.01853296160697937, + 0.018138809129595757, + 0.026985790580511093, + 0.011152039282023907, + 0.012880969792604446, + 0.008962984196841717, + 0.0075043803080916405, + 0.012367001734673977, + 0.012542215175926685, + 0.009771446697413921, + 0.021344425156712532, + 0.006755605340003967, + 0.00612784456461668, + 0.006951998453587294, + 0.012931781820952892, + 0.00814487598836422, + 0.01322534866631031, + 0.00737669225782156, + 0.007562599144876003, + 0.009033593349158764, + 0.005781993269920349, + 0.011736879125237465, + 0.017407268285751343, + 0.011409482918679714, + 0.014515696093440056, + 0.007596577983349562, + 0.005701521877199411, + 0.0029977106023579836, + 0.0024435173254460096, + 0.0033561312593519688, + 0.0038988490123301744, + 0.004533541388809681, + 0.0022814071271568537, + 0.0033206422813236713, + 0.0022234530188143253, + 0.0032965796999633312, + 0.003277750685811043, + 0.0031638285145163536, + 0.00323269865475595, + 0.0031514533329755068, + 0.002775464905425906, + 0.0032202666625380516, + 0.0026622689329087734, + 0.010598644614219666, + 0.010623359121382236, + 0.012936502695083618, + 0.012609107419848442, + 0.008570482954382896, + 0.009546690620481968, + 0.010524403303861618, + 0.006771083455532789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.006746135652065277 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002081722021103 + ], + "opposite_logprob_tokens": [ + -2.695020914077759 + ], + "logprob_reward": -0.07002081722021103, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002081722021103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1322432905435562, + 0.0762481763958931, + 0.06280086189508438, + 0.2067377269268036, + 0.30818042159080505, + 0.12037833034992218, + 0.04652278870344162, + 0.015098536387085915, + 0.01043544802814722, + 0.021354442462325096 + ], + "credit_map": [ + 0.008002467453479767, + 0.013997985981404781, + 0.01127748005092144, + 0.015467744320631027, + 0.022693971171975136, + 0.014024884440004826, + 0.020125608891248703, + 0.026653148233890533, + 0.006989744957536459, + 0.006959595717489719, + 0.007891974411904812, + 0.00656012212857604, + 0.00997775886207819, + 0.009096783585846424, + 0.012291214428842068, + 0.016480984166264534, + 0.0055641368962824345, + 0.010976857505738735, + 0.004798302426934242, + 0.011739523150026798, + 0.007358633913099766, + 0.007592122536152601, + 0.008191102184355259, + 0.006580181419849396, + 0.013797786086797714, + 0.01568414829671383, + 0.01637526974081993, + 0.02527908980846405, + 0.025182463228702545, + 0.03114992380142212, + 0.021230654790997505, + 0.058038391172885895, + 0.021473634988069534, + 0.034777093678712845, + 0.0368652269244194, + 0.030365299433469772, + 0.03912797570228577, + 0.0419103279709816, + 0.03136765956878662, + 0.07229319959878922, + 0.009514839388430119, + 0.009519154205918312, + 0.014562437310814857, + 0.013722682371735573, + 0.016909755766391754, + 0.017396947368979454, + 0.014657811261713505, + 0.024094702675938606, + 0.00474507175385952, + 0.006534751038998365, + 0.0058412509970366955, + 0.00478897662833333, + 0.009775499813258648, + 0.005481216125190258, + 0.005164801608771086, + 0.004191219806671143, + 0.0017342623323202133, + 0.0012191219720989466, + 0.0023839804343879223, + 0.0018613521242514253, + 0.0017299692844972014, + 0.0016772076487541199, + 0.0019855722784996033, + 0.00250707077793777, + 0.001310633379034698, + 0.0015741699608042836, + 0.0015221909852698445, + 0.0011820683721452951, + 0.0014855586923658848, + 0.0011387536069378257, + 0.001251041772775352, + 0.0009710309677757323, + 0.003090854734182358, + 0.0026265105698257685, + 0.0027772176545113325, + 0.0029601638671010733, + 0.0026634312234818935, + 0.0019257920794188976, + 0.003264255356043577, + 0.0020462169777601957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.07002081722021103 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the spacecraft massive with a metal hull covered in weathered titanium panels and a complex network of pipes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are massive engine nozzles at the rear of the ship belching a burst of flames with a near-incandescent white core fading to orange and yellow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a torrent of thick white smoke and steam erupting from the base, engulfing the concrete and steel launch pad?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep blue predawn sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start at an extremely low angle and tilt slowly and steadily upward following the spacecraft?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328714169561863 + ], + "opposite_logprob_tokens": [ + -3.773287057876587 + ], + "logprob_reward": -0.02328714169561863, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328714169561863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12060364335775375, + 0.09208707511425018, + 0.08541391789913177, + 0.24109046161174774, + 0.20035339891910553, + 0.09843884408473969, + 0.04400063306093216, + 0.044861532747745514, + 0.04906637594103813, + 0.02408415451645851 + ], + "credit_map": [ + 0.013809345662593842, + 0.018520541489124298, + 0.013869037851691246, + 0.0131188053637743, + 0.014515414834022522, + 0.012277651578187943, + 0.014084281399846077, + 0.02040856145322323, + 0.009208137169480324, + 0.011679778806865215, + 0.013748754747211933, + 0.009727699682116508, + 0.01157799456268549, + 0.009135746397078037, + 0.015276915393769741, + 0.01173204556107521, + 0.006946953013539314, + 0.015652313828468323, + 0.006620393600314856, + 0.013774220831692219, + 0.011074814014136791, + 0.010704364627599716, + 0.010620023123919964, + 0.010020836256444454, + 0.013902271166443825, + 0.021762972697615623, + 0.02457963116466999, + 0.06279535591602325, + 0.03189432993531227, + 0.029541637748479843, + 0.028521256521344185, + 0.028093021363019943, + 0.021237270906567574, + 0.020542442798614502, + 0.030653441324830055, + 0.021955400705337524, + 0.024365447461605072, + 0.024447860196232796, + 0.020499469712376595, + 0.03665206581354141, + 0.00725126825273037, + 0.009470184333622456, + 0.0110546313226223, + 0.012774016708135605, + 0.014729462563991547, + 0.013014898635447025, + 0.013266189023852348, + 0.016878195106983185, + 0.003893162589520216, + 0.0034578407648950815, + 0.004843148868530989, + 0.006522683426737785, + 0.007958888076245785, + 0.007301611360162497, + 0.005335337482392788, + 0.004687964450567961, + 0.006471194326877594, + 0.006268853321671486, + 0.005918371491134167, + 0.004178691189736128, + 0.007372160442173481, + 0.0067640808410942554, + 0.004263080656528473, + 0.0036250967532396317, + 0.0053141252137720585, + 0.005752606317400932, + 0.009793343022465706, + 0.005292112939059734, + 0.006086488254368305, + 0.005819152574986219, + 0.00541324820369482, + 0.005595295690000057, + 0.0029635638929903507, + 0.004025730304419994, + 0.004081191029399633, + 0.0028394705150276423, + 0.0027491964865475893, + 0.002152827102690935, + 0.0036295277532190084, + 0.0016426487127318978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.02328714169561863 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_6/index_tensor([166, 166])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_6/index_tensor([166, 166])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..31cd59935c01b305e782dfa99d2811a2e7e27fc9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_6/index_tensor([166, 166])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely close, still photograph is held flush against the surface of a dark, polished oak table. A sheet of off-white paper, crumpled into a complex network of sharp creases and soft wrinkles, balances precariously on one edge of its folded edge, forming a fragile, makeshift sculpture against the deep grain of the wood. A single, warm light source on the right side of the frame casts long, dramatic shadows across the table and deep within the paper's folds, accentuating its three-dimensional texture and creating an atmosphere of quiet, hushed tension. After a moment of perfect stillness, an almost imperceptible tremor emerges from the paper's structure; its center of gravity decisively shifts, it yields swiftly and silently, its folds collapsing in on themselves as it falls flat on the wooden surface in a soft, perfect motion, its complex form instantly reduced to a two-dimensional, crumpled sheet.", + "seed": 19414, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_6/index_166_seed_19414_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.060998429404571654, + "vlm_logprob": -0.060998429404571654, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.060998429404571654, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01783161051571369 + ], + "opposite_logprob_tokens": [ + -4.392831802368164 + ], + "logprob_reward": -0.01783161051571369, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01783161051571369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30288833379745483, + 0.08710993826389313, + 0.20824076235294342, + 0.07833579182624817, + 0.03977416455745697, + 0.05161743238568306, + 0.12759263813495636, + 0.03366154432296753, + 0.04732495918869972, + 0.023454368114471436 + ], + "credit_map": [ + 0.04813612625002861, + 0.030588405206799507, + 0.03887090086936951, + 0.03132893517613411, + 0.02615700289607048, + 0.04181404039263725, + 0.03821045905351639, + 0.04778248444199562, + 0.011476757936179638, + 0.010555576533079147, + 0.008054744452238083, + 0.011960680596530437, + 0.01060450915247202, + 0.010648414492607117, + 0.011856690049171448, + 0.011952565051615238, + 0.022940728813409805, + 0.027664849534630775, + 0.013983575627207756, + 0.03513403981924057, + 0.02761741913855076, + 0.025123663246631622, + 0.0321633405983448, + 0.023613153025507927, + 0.008068976923823357, + 0.008646714501082897, + 0.005481645464897156, + 0.01336020976305008, + 0.01229402981698513, + 0.009829284623265266, + 0.011875339783728123, + 0.00877959281206131, + 0.004297399427741766, + 0.003656218061223626, + 0.003435765393078327, + 0.005751433782279491, + 0.0045097521506249905, + 0.005430163815617561, + 0.006502282340079546, + 0.006191151682287455, + 0.00788412056863308, + 0.005052976310253143, + 0.008513915352523327, + 0.008556060492992401, + 0.005170740187168121, + 0.006807548925280571, + 0.00469279894605279, + 0.004939273931086063, + 0.014931696467101574, + 0.03414616733789444, + 0.005582209676504135, + 0.00897059217095375, + 0.021124109625816345, + 0.02900479920208454, + 0.0056837014853954315, + 0.008149362169206142, + 0.003778639016672969, + 0.007559965830296278, + 0.0018332306062802672, + 0.004022839944809675, + 0.005506014451384544, + 0.006303157191723585, + 0.0018532167887315154, + 0.002804481191560626, + 0.007366064935922623, + 0.009391659870743752, + 0.003309200517833233, + 0.005269770044833422, + 0.006726596504449844, + 0.0075769065879285336, + 0.002466559410095215, + 0.005218201782554388, + 0.0030620365869253874, + 0.005584957543760538, + 0.0016195395728573203, + 0.002853674814105034, + 0.004078260622918606, + 0.0029300639871507883, + 0.0012869976926594973, + 0.0020388360135257244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.01783161051571369 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020608466118574142 + ], + "opposite_logprob_tokens": [ + -3.895608425140381 + ], + "logprob_reward": -0.020608466118574142, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020608466118574142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2629464268684387, + 0.21604372560977936, + 0.013952594250440598, + 0.13010074198246002, + 0.04527745768427849, + 0.07113297283649445, + 0.14963005483150482, + 0.03439329192042351, + 0.0326385423541069, + 0.043884165585041046 + ], + "credit_map": [ + 0.042639341205358505, + 0.025153998285531998, + 0.033287663012742996, + 0.023670095950365067, + 0.023101527243852615, + 0.03687899932265282, + 0.03507719933986664, + 0.04313761740922928, + 0.03005693294107914, + 0.02442113868892193, + 0.02342407964169979, + 0.025183064863085747, + 0.02487478032708168, + 0.0262908898293972, + 0.02730817161500454, + 0.034484684467315674, + 0.0017719344468787313, + 0.0012918952852487564, + 0.0018467891495674849, + 0.0017719765892252326, + 0.0015899018617346883, + 0.0017511891201138496, + 0.001783148618414998, + 0.002145761623978615, + 0.013702142983675003, + 0.014627164229750633, + 0.009512471966445446, + 0.021274153143167496, + 0.020347563549876213, + 0.016728103160858154, + 0.01847839169204235, + 0.015430759638547897, + 0.004730632528662682, + 0.004388424102216959, + 0.004384566564112902, + 0.005803581792861223, + 0.006648534908890724, + 0.005308129824697971, + 0.007021098397672176, + 0.006992492824792862, + 0.010520536452531815, + 0.011046633124351501, + 0.011992113664746284, + 0.011724928393959999, + 0.006579320877790451, + 0.006950670853257179, + 0.006091793533414602, + 0.0062269833870232105, + 0.01991201378405094, + 0.040803007781505585, + 0.004868902266025543, + 0.010969123803079128, + 0.024205194786190987, + 0.033773425966501236, + 0.005588812287896872, + 0.009509585797786713, + 0.0037320174742490053, + 0.007558304816484451, + 0.0019035309087485075, + 0.0036605577915906906, + 0.005590920802205801, + 0.00666609313338995, + 0.0018637761240825057, + 0.003418095177039504, + 0.004408213775604963, + 0.007769672200083733, + 0.0014694900019094348, + 0.0027912992518395185, + 0.0048707216046750546, + 0.006707982625812292, + 0.0014189350185915828, + 0.003202234860509634, + 0.005624013487249613, + 0.012361759319901466, + 0.0014542380813509226, + 0.002785698976367712, + 0.007360171526670456, + 0.010022353380918503, + 0.0017033235635608435, + 0.0025726130697876215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.020608466118574142 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562137246131897 + ], + "opposite_logprob_tokens": [ + -1.6006213426589966 + ], + "logprob_reward": -0.22562137246131897, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562137246131897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2531087100505829, + 0.20575611293315887, + 0.06838640570640564, + 0.20809590816497803, + 0.06019170209765434, + 0.05639290437102318, + 0.06367026269435883, + 0.04784631356596947, + 0.019169578328728676, + 0.017382098361849785 + ], + "credit_map": [ + 0.037828680127859116, + 0.020806556567549706, + 0.03112448938190937, + 0.0261568371206522, + 0.023135840892791748, + 0.03348138928413391, + 0.03290240094065666, + 0.047672513872385025, + 0.02775338664650917, + 0.020239543169736862, + 0.017856424674391747, + 0.02053048089146614, + 0.022960327565670013, + 0.024151602759957314, + 0.03743145987391472, + 0.0348329059779644, + 0.006346500013023615, + 0.008211717009544373, + 0.005324509460479021, + 0.011386781930923462, + 0.008237145841121674, + 0.007364747114479542, + 0.011137688532471657, + 0.010377313010394573, + 0.02171768993139267, + 0.023236842826008797, + 0.014550954103469849, + 0.0375436469912529, + 0.029113460332155228, + 0.02368876338005066, + 0.033353399485349655, + 0.024891149252653122, + 0.006397060584276915, + 0.005252963397651911, + 0.005430057644844055, + 0.007946670986711979, + 0.0070149837993085384, + 0.006889040116220713, + 0.010245932266116142, + 0.011014995165169239, + 0.007900307886302471, + 0.008810634724795818, + 0.008118304423987865, + 0.006146226543933153, + 0.00675737950950861, + 0.008105304092168808, + 0.006458699237555265, + 0.004096049815416336, + 0.009027752093970776, + 0.016881443560123444, + 0.002693825401365757, + 0.004833537619560957, + 0.010101147927343845, + 0.012597735971212387, + 0.0025452133268117905, + 0.004989602603018284, + 0.006106959655880928, + 0.011493728496134281, + 0.0027607691008597612, + 0.004380857106298208, + 0.00720391096547246, + 0.010153166018426418, + 0.002331632887944579, + 0.0034152904991060495, + 0.002511797472834587, + 0.0038681209553033113, + 0.0013718794798478484, + 0.0020963873248547316, + 0.00267576496116817, + 0.003393217921257019, + 0.0014211783418431878, + 0.0018312330357730389, + 0.0021770442835986614, + 0.003857482923194766, + 0.0008794972090981901, + 0.0020730511751025915, + 0.003014662768691778, + 0.0034309453330934048, + 0.0008527772151865065, + 0.0010966366389766335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.22562137246131897 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811061009764671 + ], + "opposite_logprob_tokens": [ + -3.2881104946136475 + ], + "logprob_reward": -0.03811061009764671, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811061009764671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09014632552862167, + 0.0947134718298912, + 0.1209811195731163, + 0.16664771735668182, + 0.058090124279260635, + 0.20417743921279907, + 0.13645489513874054, + 0.033784814178943634, + 0.03352382779121399, + 0.06148028001189232 + ], + "credit_map": [ + 0.015016520395874977, + 0.01119143608957529, + 0.009085249155759811, + 0.01095461193472147, + 0.00898663979023695, + 0.010537405498325825, + 0.013983299024403095, + 0.010391172021627426, + 0.012889105826616287, + 0.012033559381961823, + 0.011279375292360783, + 0.010839609429240227, + 0.010841287672519684, + 0.010923297144472599, + 0.010414293967187405, + 0.015492941252887249, + 0.014662363566458225, + 0.01555465254932642, + 0.008516928181052208, + 0.020145822316408157, + 0.016019342467188835, + 0.014155600219964981, + 0.01868261769413948, + 0.013243790715932846, + 0.018555961549282074, + 0.018615923821926117, + 0.012426158413290977, + 0.02864186465740204, + 0.02477177418768406, + 0.020070524886250496, + 0.023473581299185753, + 0.020091937854886055, + 0.006380031816661358, + 0.005570304114371538, + 0.005810783710330725, + 0.00750755425542593, + 0.007004016079008579, + 0.0070359427481889725, + 0.009020663797855377, + 0.009760822169482708, + 0.03070804663002491, + 0.029287662357091904, + 0.030356597155332565, + 0.02710017003118992, + 0.02069171890616417, + 0.027300048619508743, + 0.018188005313277245, + 0.020545199513435364, + 0.019766511395573616, + 0.031710490584373474, + 0.006994164548814297, + 0.011106963269412518, + 0.021071162074804306, + 0.02998051978647709, + 0.005949489772319794, + 0.009875589050352573, + 0.004289382603019476, + 0.00740460492670536, + 0.0026568383909761906, + 0.003964591771364212, + 0.00548902153968811, + 0.005442441441118717, + 0.0016057563479989767, + 0.002932177158072591, + 0.004348088055849075, + 0.008178102783858776, + 0.002322869375348091, + 0.002659542951732874, + 0.006141546182334423, + 0.005337058566510677, + 0.0016253892099484801, + 0.002911231480538845, + 0.007554417010396719, + 0.01735106110572815, + 0.002075823489576578, + 0.004248965065926313, + 0.010201048105955124, + 0.013886637054383755, + 0.0024861323181539774, + 0.0036761986557394266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.03811061009764671 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820087829604745 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820087829604745, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820087829604745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3155377507209778, + 0.061552632600069046, + 0.12658002972602844, + 0.07382051646709442, + 0.05753426253795624, + 0.12833864986896515, + 0.11789407581090927, + 0.0456254668533802, + 0.0328386053442955, + 0.04027806222438812 + ], + "credit_map": [ + 0.05064534395933151, + 0.03162551671266556, + 0.0396585650742054, + 0.027521569281816483, + 0.027558593079447746, + 0.043085068464279175, + 0.04605625942349434, + 0.049386803060770035, + 0.009244753047823906, + 0.007741613779217005, + 0.0057267420925199986, + 0.006500269751995802, + 0.007226777262985706, + 0.007731377612799406, + 0.008510097861289978, + 0.008870997466146946, + 0.013498195447027683, + 0.016651194542646408, + 0.008752790279686451, + 0.021631840616464615, + 0.01661083661019802, + 0.014284959062933922, + 0.019740892574191093, + 0.015409305691719055, + 0.007476524915546179, + 0.007690139580518007, + 0.005733211990445852, + 0.012208310887217522, + 0.010701995342969894, + 0.007975040934979916, + 0.012358633801341057, + 0.009676648303866386, + 0.006434768438339233, + 0.005885239690542221, + 0.005494516808539629, + 0.00697886710986495, + 0.007507225964218378, + 0.0071214050985872746, + 0.008509010076522827, + 0.009603225626051426, + 0.0192632544785738, + 0.015176971442997456, + 0.020052339881658554, + 0.015155741013586521, + 0.012895351275801659, + 0.018780551850795746, + 0.01307640504091978, + 0.013938038609921932, + 0.015047878958284855, + 0.022608254104852676, + 0.011226383037865162, + 0.01411829050630331, + 0.018709296360611916, + 0.018762752413749695, + 0.008184783160686493, + 0.009236421436071396, + 0.005179570522159338, + 0.010011935606598854, + 0.003299851668998599, + 0.006371929310262203, + 0.007082889787852764, + 0.0070464531891047955, + 0.0028115080203861, + 0.003821325022727251, + 0.004773762077093124, + 0.006030457094311714, + 0.002374748233705759, + 0.0037783042062073946, + 0.004549495875835419, + 0.006395271979272366, + 0.0018847010796889663, + 0.003051862120628357, + 0.005251020193099976, + 0.008278632536530495, + 0.0026667676866054535, + 0.004836694337427616, + 0.005814520176500082, + 0.007824026979506016, + 0.00230706506408751, + 0.0032993326894938946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.002820087829604745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close, still photograph is held flush against the surface of a dark, polished oak table. A sheet of off-white paper, crumpled into a complex network of sharp creases and soft wrinkles, balances precariously on one edge of its folded edge, forming a fragile, makeshift sculpture against the deep grain of the wood. A single, warm light source on the right side of the frame casts long, dramatic shadows across the table and deep within the paper's folds, accentuating its three-dimensional texture and creating an atmosphere of quiet, hushed tension. After a moment of perfect stillness, an almost imperceptible tremor emerges from the paper's structure; its center of gravity decisively shifts, it yields swiftly and silently, its folds collapsing in on themselves as it falls flat on the wooden surface in a soft, perfect motion, its complex form instantly reduced to a two-dimensional, crumpled sheet.", + "seed": 19415, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_6/index_166_seed_19415_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.030926145566627383, + "vlm_logprob": -0.030926145566627383, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.030926145566627383, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03657129779458046 + ], + "opposite_logprob_tokens": [ + -3.5365712642669678 + ], + "logprob_reward": -0.03657129779458046, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03657129779458046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02916540578007698, + 0.060281902551651, + 0.039362289011478424, + 0.041437532752752304, + 0.029454926028847694, + 0.38321158289909363, + 0.1189783588051796, + 0.10712505877017975, + 0.14672666788101196, + 0.04425631836056709 + ], + "credit_map": [ + 0.0028134251479059458, + 0.0025243565905839205, + 0.005248531233519316, + 0.0033431530464440584, + 0.00412308145314455, + 0.0046128686517477036, + 0.0025454009883105755, + 0.003954591229557991, + 0.005012300331145525, + 0.004600390326231718, + 0.010270561091601849, + 0.004044058732688427, + 0.010595841333270073, + 0.00557192787528038, + 0.005113034974783659, + 0.01507379300892353, + 0.0047768191434443, + 0.0017938989913091063, + 0.013064436614513397, + 0.0032775152940303087, + 0.00206454424187541, + 0.002617001300677657, + 0.00263552600517869, + 0.00913255289196968, + 0.005043508019298315, + 0.0030058936681598425, + 0.009115406312048435, + 0.0034143561497330666, + 0.004233347252011299, + 0.0042902217246592045, + 0.004083291627466679, + 0.008251511491835117, + 0.0036991324741393328, + 0.002134026261046529, + 0.004644266329705715, + 0.004019309300929308, + 0.0035955659113824368, + 0.0020686157513409853, + 0.00498714204877615, + 0.004306871443986893, + 0.07707482576370239, + 0.0320422425866127, + 0.09819066524505615, + 0.04312673211097717, + 0.037974074482917786, + 0.02626284398138523, + 0.029690008610486984, + 0.03885020315647125, + 0.024127351120114326, + 0.008558549918234348, + 0.02578405663371086, + 0.010636457242071629, + 0.013533849269151688, + 0.011090060696005821, + 0.01018819771707058, + 0.015059848316013813, + 0.029607567936182022, + 0.006308615207672119, + 0.03449656814336777, + 0.012317980639636517, + 0.008041701279580593, + 0.0064653013832867146, + 0.004492967389523983, + 0.005394360050559044, + 0.048027075827121735, + 0.009197446517646313, + 0.05411681905388832, + 0.013701532036066055, + 0.006415572948753834, + 0.005759553052484989, + 0.004741832613945007, + 0.00476683909073472, + 0.007657237350940704, + 0.004058758728206158, + 0.011771122924983501, + 0.006408395245671272, + 0.002170899184420705, + 0.0025817332789301872, + 0.0036712326109409332, + 0.005936938337981701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.03657129779458046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017363247461616993 + ], + "opposite_logprob_tokens": [ + -6.376736164093018 + ], + "logprob_reward": -0.0017363247461616993, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017363247461616993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.081468865275383, + 0.053516168147325516, + 0.032014764845371246, + 0.08550164103507996, + 0.02331439219415188, + 0.13046348094940186, + 0.10469614714384079, + 0.2726406455039978, + 0.18644313514232635, + 0.029940716922283173 + ], + "credit_map": [ + 0.00782174151390791, + 0.007676747161895037, + 0.01127192098647356, + 0.01143552828580141, + 0.013591253198683262, + 0.010123670101165771, + 0.010387961752712727, + 0.009160054847598076, + 0.004905895330011845, + 0.005776761565357447, + 0.008914205245673656, + 0.005748618859797716, + 0.007528075482696295, + 0.005577824078500271, + 0.006077486090362072, + 0.008987304754555225, + 0.00549503555521369, + 0.0029920588713139296, + 0.006616578903049231, + 0.003179484512656927, + 0.003025513608008623, + 0.002920543309301138, + 0.0031396150588989258, + 0.004645938519388437, + 0.01025899313390255, + 0.007633274421095848, + 0.01696676015853882, + 0.009636366739869118, + 0.011249829083681107, + 0.009661190211772919, + 0.009735915809869766, + 0.010359321720898151, + 0.003273166250437498, + 0.0023716541472822428, + 0.00375219713896513, + 0.0020881923846900463, + 0.0035305810160934925, + 0.003347616409882903, + 0.0025170249864459038, + 0.002433961257338524, + 0.01628558710217476, + 0.014112320728600025, + 0.021756576374173164, + 0.015260721556842327, + 0.019395118579268456, + 0.012641203589737415, + 0.01422366127371788, + 0.016788311302661896, + 0.02612455189228058, + 0.006358432583510876, + 0.028886111453175545, + 0.009434220381081104, + 0.008163632825016975, + 0.00963994488120079, + 0.007467779330909252, + 0.008621486835181713, + 0.0841037854552269, + 0.012731589376926422, + 0.09395130723714828, + 0.02889074757695198, + 0.013586523942649364, + 0.013945969752967358, + 0.009439300745725632, + 0.01599145494401455, + 0.06312593072652817, + 0.009575960226356983, + 0.07179979234933853, + 0.015593908727169037, + 0.008539221249520779, + 0.006662091240286827, + 0.004687429405748844, + 0.006458835210651159, + 0.004173194989562035, + 0.003327734535560012, + 0.006363547872751951, + 0.004560521803796291, + 0.001921359566040337, + 0.0024920909199863672, + 0.0027047223411500454, + 0.004397544544190168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.0017363247461616993 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308658093214035 + ], + "opposite_logprob_tokens": [ + -3.168086528778076 + ], + "logprob_reward": -0.04308658093214035, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308658093214035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03961927816271782, + 0.03800018131732941, + 0.019006604328751564, + 0.02777409367263317, + 0.018595408648252487, + 0.06979212909936905, + 0.06059487536549568, + 0.05194007232785225, + 0.6266167163848877, + 0.04806066304445267 + ], + "credit_map": [ + 0.0041551473550498486, + 0.003765575820580125, + 0.005463824141770601, + 0.005200564861297607, + 0.005589024163782597, + 0.005117335356771946, + 0.004715465474873781, + 0.005612344481050968, + 0.0037273885682225227, + 0.0017055351054295897, + 0.007879900746047497, + 0.0030212646815925837, + 0.003831861075013876, + 0.0031464521307498217, + 0.00278598815202713, + 0.011901792138814926, + 0.0019056996097788215, + 0.0012773658381775022, + 0.0034896431025117636, + 0.002457199152559042, + 0.0022396333515644073, + 0.0023045113775879145, + 0.0030373407062143087, + 0.0022952137514948845, + 0.004282491747289896, + 0.0014073617057874799, + 0.008311078883707523, + 0.0025332868099212646, + 0.0019518786575645208, + 0.001780621474608779, + 0.002005763351917267, + 0.005501613952219486, + 0.0025484359357506037, + 0.001565865008160472, + 0.0028364602476358414, + 0.0020883961115032434, + 0.00209026993252337, + 0.002168119652196765, + 0.0030339874792844057, + 0.0022638754453510046, + 0.008899137377738953, + 0.007234685588628054, + 0.012767083011567593, + 0.007656194269657135, + 0.0106846634298563, + 0.006518720183521509, + 0.008135059848427773, + 0.00789659097790718, + 0.019870614632964134, + 0.002252730308100581, + 0.024171750992536545, + 0.005116487853229046, + 0.0020341372583061457, + 0.0024696497712284327, + 0.002042398788034916, + 0.0026371129788458347, + 0.01370516698807478, + 0.0030414832290261984, + 0.01711328886449337, + 0.00539796706289053, + 0.003844336373731494, + 0.003513340139761567, + 0.003072781953960657, + 0.00225171260535717, + 0.22999003529548645, + 0.024665292352437973, + 0.2539834678173065, + 0.05391672998666763, + 0.019246868789196014, + 0.0176638625562191, + 0.009230710566043854, + 0.017919771373271942, + 0.006760952994227409, + 0.006191743537783623, + 0.010052177123725414, + 0.007390159647911787, + 0.0024286701809614897, + 0.0027122136671096087, + 0.004319726023823023, + 0.008205024525523186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.04308658093214035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003848999738693 + ], + "opposite_logprob_tokens": [ + -2.6950385570526123 + ], + "logprob_reward": -0.07003848999738693, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003848999738693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030205480754375458, + 0.02199604921042919, + 0.04922567307949066, + 0.033281996846199036, + 0.01582326926290989, + 0.058142680674791336, + 0.23966902494430542, + 0.11429627984762192, + 0.3910011053085327, + 0.04635840654373169 + ], + "credit_map": [ + 0.003058522706851363, + 0.0033886616583913565, + 0.0030824006535112858, + 0.003816511482000351, + 0.005173556040972471, + 0.004246762488037348, + 0.003791260300204158, + 0.0036478075198829174, + 0.002065724926069379, + 0.0025349718052893877, + 0.003499720012769103, + 0.002571442862972617, + 0.0030970536172389984, + 0.0023092397022992373, + 0.0034099461045116186, + 0.0025079515762627125, + 0.006521961651742458, + 0.002523183124139905, + 0.012896825559437275, + 0.00619100546464324, + 0.0030777158681303263, + 0.0037091022823005915, + 0.006656460929661989, + 0.007649414706975222, + 0.0044455984607338905, + 0.002719770185649395, + 0.006971362512558699, + 0.0032684251200407743, + 0.004518839064985514, + 0.003258107230067253, + 0.0037035353016108274, + 0.004396360367536545, + 0.002404280938208103, + 0.001395449391566217, + 0.0026594004593789577, + 0.0022422848269343376, + 0.0018295029876753688, + 0.0016026655212044716, + 0.0017525064758956432, + 0.0019371791277080774, + 0.008818035945296288, + 0.004897186998277903, + 0.011466972529888153, + 0.0074466741643846035, + 0.007728172931820154, + 0.0047195651568472385, + 0.006110767368227243, + 0.006955304648727179, + 0.0773787871003151, + 0.008052743040025234, + 0.0999736562371254, + 0.020659111440181732, + 0.0074038319289684296, + 0.009986309334635735, + 0.00754313450306654, + 0.008671458810567856, + 0.02368916943669319, + 0.009513609111309052, + 0.03086441569030285, + 0.013168131932616234, + 0.010970566421747208, + 0.009194448590278625, + 0.010025186464190483, + 0.006870758254081011, + 0.14311560988426208, + 0.01566084288060665, + 0.15825501084327698, + 0.03315553814172745, + 0.012628943659365177, + 0.011011619120836258, + 0.005848603323101997, + 0.011324942111968994, + 0.007000565063208342, + 0.005580877885222435, + 0.005789316724985838, + 0.005876879207789898, + 0.005268950946629047, + 0.004674399271607399, + 0.006460803095251322, + 0.005706614814698696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.07003848999738693 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031980343628674746 + ], + "opposite_logprob_tokens": [ + -5.753198146820068 + ], + "logprob_reward": -0.0031980343628674746, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031980343628674746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0531177781522274, + 0.045693449676036835, + 0.05110618472099304, + 0.04083917662501335, + 0.030897580087184906, + 0.2699008584022522, + 0.29665207862854004, + 0.12297995388507843, + 0.0705166682600975, + 0.018296213820576668 + ], + "credit_map": [ + 0.0054520731791853905, + 0.005594450980424881, + 0.006894441321492195, + 0.0074137113988399506, + 0.007632855791598558, + 0.007301325909793377, + 0.0062297433614730835, + 0.006599180866032839, + 0.00362550956197083, + 0.003556248964741826, + 0.009642932564020157, + 0.00482894154265523, + 0.004839814268052578, + 0.004389411304146051, + 0.004629143979400396, + 0.010181449353694916, + 0.006414115894585848, + 0.0032694218680262566, + 0.01318669505417347, + 0.005673835054039955, + 0.003795876633375883, + 0.004119533579796553, + 0.00617316784337163, + 0.008473539724946022, + 0.005801198538392782, + 0.003293718211352825, + 0.009411520324647427, + 0.0038266933988779783, + 0.004814525600522757, + 0.0038195245433598757, + 0.00427527679130435, + 0.0055967215448617935, + 0.0037737360689789057, + 0.0031336480751633644, + 0.0041404892690479755, + 0.004265207331627607, + 0.004853273741900921, + 0.003721035085618496, + 0.003579925512894988, + 0.0034302668645977974, + 0.06463046371936798, + 0.019707895815372467, + 0.07875832915306091, + 0.029210170730948448, + 0.019254107028245926, + 0.015437841415405273, + 0.017511911690235138, + 0.0253901444375515, + 0.09138572961091995, + 0.010742761194705963, + 0.11665245145559311, + 0.024945618584752083, + 0.012980855070054531, + 0.015387766063213348, + 0.011465155519545078, + 0.013091759756207466, + 0.0275823175907135, + 0.010433923453092575, + 0.030719809234142303, + 0.01435556448996067, + 0.012416752055287361, + 0.009154566563665867, + 0.010457830503582954, + 0.007859209552407265, + 0.023285282775759697, + 0.006944385357201099, + 0.0226574819535017, + 0.005461765918880701, + 0.0032745215576142073, + 0.0037036926951259375, + 0.002592681208625436, + 0.0025968661066144705, + 0.003337116679176688, + 0.0018437649123370647, + 0.0027157950680702925, + 0.00286882184445858, + 0.001997451763600111, + 0.0013132636668160558, + 0.0020305467769503593, + 0.0021894541569054127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.0031980343628674746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close, still photograph is held flush against the surface of a dark, polished oak table. A sheet of off-white paper, crumpled into a complex network of sharp creases and soft wrinkles, balances precariously on one edge of its folded edge, forming a fragile, makeshift sculpture against the deep grain of the wood. A single, warm light source on the right side of the frame casts long, dramatic shadows across the table and deep within the paper's folds, accentuating its three-dimensional texture and creating an atmosphere of quiet, hushed tension. After a moment of perfect stillness, an almost imperceptible tremor emerges from the paper's structure; its center of gravity decisively shifts, it yields swiftly and silently, its folds collapsing in on themselves as it falls flat on the wooden surface in a soft, perfect motion, its complex form instantly reduced to a two-dimensional, crumpled sheet.", + "seed": 19416, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_6/index_166_seed_19416_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02898054742254317, + "vlm_logprob": -0.02898054742254317, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02898054742254317, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03309488296508789 + ], + "opposite_logprob_tokens": [ + -3.783094882965088 + ], + "logprob_reward": -0.03309488296508789, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03309488296508789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10759853571653366, + 0.052294500172138214, + 0.06202284246683121, + 0.06045638769865036, + 0.16765865683555603, + 0.1493198424577713, + 0.07714275270700455, + 0.20558835566043854, + 0.05544449761509895, + 0.062473636120557785 + ], + "credit_map": [ + 0.027836663648486137, + 0.01151860598474741, + 0.010561679489910603, + 0.00842193141579628, + 0.011383491568267345, + 0.013688034377992153, + 0.007206459064036608, + 0.01698167249560356, + 0.0057768248952925205, + 0.008229040540754795, + 0.0043041035532951355, + 0.0071028065867722034, + 0.007584829349070787, + 0.007195210549980402, + 0.006183760706335306, + 0.005917925853282213, + 0.012982914224267006, + 0.007798635866492987, + 0.005770491901785135, + 0.006396447774022818, + 0.007456356659531593, + 0.0066762883216142654, + 0.0048995912075042725, + 0.010042115114629269, + 0.01549433171749115, + 0.007046619895845652, + 0.004249458201229572, + 0.004923240281641483, + 0.006485207937657833, + 0.007062081713229418, + 0.005761300679296255, + 0.009434151463210583, + 0.034195348620414734, + 0.022560512647032738, + 0.013431058265268803, + 0.015461016446352005, + 0.026021959260106087, + 0.02463673986494541, + 0.01238186378031969, + 0.01897016353905201, + 0.026066159829497337, + 0.022990068420767784, + 0.014529717154800892, + 0.012206118553876877, + 0.027313604950904846, + 0.018477044999599457, + 0.012610779143869877, + 0.015126350335776806, + 0.011435270309448242, + 0.013978170230984688, + 0.007545944768935442, + 0.006382908206433058, + 0.012113450095057487, + 0.012091664597392082, + 0.005256777163594961, + 0.008338567800819874, + 0.027679959312081337, + 0.03206603601574898, + 0.020674094557762146, + 0.015990130603313446, + 0.033717330545186996, + 0.029780825600028038, + 0.01958751678466797, + 0.026092471554875374, + 0.007820899598300457, + 0.008388962596654892, + 0.00807503517717123, + 0.004722477402538061, + 0.008879544213414192, + 0.0070427111349999905, + 0.0037022209726274014, + 0.006812647450715303, + 0.010212909430265427, + 0.009385970421135426, + 0.00660961028188467, + 0.004670984577387571, + 0.010019557550549507, + 0.010049143806099892, + 0.003904994111508131, + 0.007620468735694885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.03309488296508789 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032202552538365126 + ], + "opposite_logprob_tokens": [ + -5.753220081329346 + ], + "logprob_reward": -0.0032202552538365126, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032202552538365126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28857091069221497, + 0.08718662708997726, + 0.04654315486550331, + 0.08831338584423065, + 0.1045338362455368, + 0.07549160718917847, + 0.06290598958730698, + 0.055792707949876785, + 0.15334127843379974, + 0.03732047602534294 + ], + "credit_map": [ + 0.0490359328687191, + 0.04497772827744484, + 0.026169126853346825, + 0.03626172989606857, + 0.044396329671144485, + 0.02899809367954731, + 0.023274224251508713, + 0.03545774146914482, + 0.016899077221751213, + 0.013471437618136406, + 0.006107161287218332, + 0.008294159546494484, + 0.011887808330357075, + 0.011043623089790344, + 0.008368222042918205, + 0.01111513376235962, + 0.005874693859368563, + 0.006793492008000612, + 0.004339613486081362, + 0.005541207268834114, + 0.006814152933657169, + 0.005030706990510225, + 0.006242027506232262, + 0.005907262675464153, + 0.016757668927311897, + 0.014359376393258572, + 0.005672903265804052, + 0.006608856376260519, + 0.015923071652650833, + 0.010220388881862164, + 0.008352428674697876, + 0.010418694466352463, + 0.026369595900177956, + 0.01208052784204483, + 0.007087542675435543, + 0.007957442663609982, + 0.012924596667289734, + 0.01477658562362194, + 0.00792086310684681, + 0.015416687354445457, + 0.013074290007352829, + 0.012782645411789417, + 0.0048737358301877975, + 0.0055953869596123695, + 0.015536535531282425, + 0.009846769273281097, + 0.006728954613208771, + 0.0070532867684960365, + 0.008912491612136364, + 0.012796401977539062, + 0.004776609130203724, + 0.004643039312213659, + 0.013084936887025833, + 0.007053919602185488, + 0.005306145176291466, + 0.006332442630082369, + 0.00859660655260086, + 0.007936558686196804, + 0.004650656599551439, + 0.004292488098144531, + 0.01067241933196783, + 0.008558517321944237, + 0.004234289284795523, + 0.006851173937320709, + 0.022791121155023575, + 0.02634049579501152, + 0.013468150049448013, + 0.014424500986933708, + 0.03282574564218521, + 0.01732807233929634, + 0.010707685723900795, + 0.015455508604645729, + 0.009591083973646164, + 0.0047765206545591354, + 0.0026796276215463877, + 0.0026852425653487444, + 0.005679801106452942, + 0.005044597666710615, + 0.0024572876282036304, + 0.004406314808875322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0032202552538365126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902964949607849 + ], + "opposite_logprob_tokens": [ + -2.5790295600891113 + ], + "logprob_reward": -0.07902964949607849, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902964949607849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23468321561813354, + 0.29029378294944763, + 0.031443025916814804, + 0.027648067101836205, + 0.06063925102353096, + 0.0628112256526947, + 0.05083354562520981, + 0.11312030255794525, + 0.05609669163823128, + 0.07243090867996216 + ], + "credit_map": [ + 0.028822487220168114, + 0.029402634128928185, + 0.028606947511434555, + 0.02651130221784115, + 0.030031075701117516, + 0.03325968235731125, + 0.023015262559056282, + 0.03503379598259926, + 0.05699072405695915, + 0.0243313480168581, + 0.029366495087742805, + 0.022496363148093224, + 0.028420593589544296, + 0.05386783182621002, + 0.02400529943406582, + 0.0508151538670063, + 0.007536440156400204, + 0.003932929132133722, + 0.002605790039524436, + 0.0028624346014112234, + 0.004056702833622694, + 0.00397960189729929, + 0.002247828058898449, + 0.00422130199149251, + 0.0038789266254752874, + 0.0027213185094296932, + 0.0029971550684422255, + 0.0030523091554641724, + 0.0035371962003409863, + 0.004640784580260515, + 0.0030194607097655535, + 0.0038009178824722767, + 0.009738398715853691, + 0.0101574482396245, + 0.004496105946600437, + 0.006097974721342325, + 0.010571666993200779, + 0.008531433530151844, + 0.004845785442739725, + 0.006200434174388647, + 0.00993184931576252, + 0.00993437971919775, + 0.00550089729949832, + 0.006004470866173506, + 0.010291818529367447, + 0.009147671051323414, + 0.00578682916238904, + 0.006213306915014982, + 0.010883482173085213, + 0.006045844405889511, + 0.0067399851977825165, + 0.0046153380535542965, + 0.005566244013607502, + 0.007127238903194666, + 0.004486656282097101, + 0.0053687565959990025, + 0.015919925644993782, + 0.01662958785891533, + 0.011917867697775364, + 0.009329289197921753, + 0.016711901873350143, + 0.015980999916791916, + 0.011635330505669117, + 0.014995408244431019, + 0.00917819980531931, + 0.0073062046431005, + 0.006182712968438864, + 0.0047379471361637115, + 0.012260736897587776, + 0.006250778213143349, + 0.004472392611205578, + 0.0057077263481915, + 0.014623941853642464, + 0.009261508472263813, + 0.007626341190189123, + 0.00498714717105031, + 0.009244193322956562, + 0.012986704707145691, + 0.004302839282900095, + 0.009398229420185089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.07902964949607849 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026359083130955696 + ], + "opposite_logprob_tokens": [ + -3.6513590812683105 + ], + "logprob_reward": -0.026359083130955696, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026359083130955696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15524035692214966, + 0.09039265662431717, + 0.04656635969877243, + 0.10361967980861664, + 0.15907631814479828, + 0.040397658944129944, + 0.067420095205307, + 0.1244443878531456, + 0.18438664078712463, + 0.028455782681703568 + ], + "credit_map": [ + 0.047021277248859406, + 0.014525429345667362, + 0.008622203022241592, + 0.01388796791434288, + 0.014224560000002384, + 0.0178304985165596, + 0.009619612246751785, + 0.029508834704756737, + 0.025573786348104477, + 0.009655105881392956, + 0.005150096490979195, + 0.006923717446625233, + 0.009534941054880619, + 0.013299214653670788, + 0.004828588105738163, + 0.01542721875011921, + 0.009234572760760784, + 0.006107627879828215, + 0.004225323908030987, + 0.00474211759865284, + 0.006449836306273937, + 0.005019026342779398, + 0.004455138463526964, + 0.006332725286483765, + 0.01122205425053835, + 0.013708997517824173, + 0.013757690787315369, + 0.012552149593830109, + 0.013338903896510601, + 0.01381007581949234, + 0.012579490430653095, + 0.012650340795516968, + 0.04480395093560219, + 0.015863865613937378, + 0.01049031037837267, + 0.010184936225414276, + 0.016847239807248116, + 0.023549679666757584, + 0.011785543523728848, + 0.02555081807076931, + 0.010469789616763592, + 0.004608831834048033, + 0.00416709017008543, + 0.0031488805543631315, + 0.0044525396078825, + 0.005943266209214926, + 0.002709620399400592, + 0.004897648468613625, + 0.014499547891318798, + 0.005181893706321716, + 0.009729734621942043, + 0.00680017052218318, + 0.006535637192428112, + 0.010317311622202396, + 0.006024862173944712, + 0.00833095982670784, + 0.017054766416549683, + 0.016880933195352554, + 0.012524206191301346, + 0.012167219072580338, + 0.01620454527437687, + 0.01791451685130596, + 0.015042606741189957, + 0.016655612736940384, + 0.029420480132102966, + 0.032974306493997574, + 0.01621357724070549, + 0.01579681597650051, + 0.036664918065071106, + 0.020631438121199608, + 0.012935358099639416, + 0.019749779254198074, + 0.005948699079453945, + 0.005575788673013449, + 0.002313393633812666, + 0.0027711617294698954, + 0.0034703596029430628, + 0.0034153193701058626, + 0.001806193613447249, + 0.0031548715196549892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.026359083130955696 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319886626675725 + ], + "opposite_logprob_tokens": [ + -5.753199100494385 + ], + "logprob_reward": -0.00319886626675725, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319886626675725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16831935942173004, + 0.11025112122297287, + 0.07976485043764114, + 0.04552248865365982, + 0.08678911626338959, + 0.06815270334482193, + 0.06200847774744034, + 0.09311899542808533, + 0.19193893671035767, + 0.09413392841815948 + ], + "credit_map": [ + 0.04489778354763985, + 0.01853138767182827, + 0.010257086716592312, + 0.018390106037259102, + 0.019594179466366768, + 0.016663379967212677, + 0.010256496258080006, + 0.029728936031460762, + 0.032067738473415375, + 0.01408236101269722, + 0.005857008043676615, + 0.006687814369797707, + 0.009060532785952091, + 0.01570834405720234, + 0.006245019380003214, + 0.02054229937493801, + 0.021234899759292603, + 0.008219120092689991, + 0.004675684496760368, + 0.006545949261635542, + 0.007104156073182821, + 0.009879181161522865, + 0.007100420538336039, + 0.015005437657237053, + 0.006792206317186356, + 0.005280311685055494, + 0.005439489148557186, + 0.004159554839134216, + 0.006475082132965326, + 0.006777232047170401, + 0.004749697167426348, + 0.005848917178809643, + 0.025830309838056564, + 0.009129821322858334, + 0.004738728050142527, + 0.005219494923949242, + 0.0077462769113481045, + 0.012071776203811169, + 0.006589515600353479, + 0.015463191084563732, + 0.010873442515730858, + 0.012087793089449406, + 0.005523766856640577, + 0.006123572122305632, + 0.011478139087557793, + 0.009107363410294056, + 0.00674402853474021, + 0.006214592140167952, + 0.01180871482938528, + 0.007896565832197666, + 0.007345816120505333, + 0.00542161799967289, + 0.00871428195387125, + 0.008599484339356422, + 0.005160951986908913, + 0.0070610446855425835, + 0.012724008411169052, + 0.013760773465037346, + 0.008298241533339024, + 0.006904517766088247, + 0.015286531299352646, + 0.015057557262480259, + 0.008651906624436378, + 0.01243546325713396, + 0.03178798779845238, + 0.031282469630241394, + 0.015965426340699196, + 0.017763007432222366, + 0.03518016263842583, + 0.024267079308629036, + 0.014420284889638424, + 0.02127252332866192, + 0.016447262838482857, + 0.014437169767916203, + 0.007994932122528553, + 0.009872722439467907, + 0.016219235956668854, + 0.012291737832129002, + 0.007364021148532629, + 0.009506846778094769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.00319886626675725 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely close, still photograph is held flush against the surface of a dark, polished oak table. A sheet of off-white paper, crumpled into a complex network of sharp creases and soft wrinkles, balances precariously on one edge of its folded edge, forming a fragile, makeshift sculpture against the deep grain of the wood. A single, warm light source on the right side of the frame casts long, dramatic shadows across the table and deep within the paper's folds, accentuating its three-dimensional texture and creating an atmosphere of quiet, hushed tension. After a moment of perfect stillness, an almost imperceptible tremor emerges from the paper's structure; its center of gravity decisively shifts, it yields swiftly and silently, its folds collapsing in on themselves as it falls flat on the wooden surface in a soft, perfect motion, its complex form instantly reduced to a two-dimensional, crumpled sheet.", + "seed": 19417, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_6/index_166_seed_19417_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018122045043855905, + "vlm_logprob": -0.018122045043855905, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.018122045043855905, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02285834215581417 + ], + "opposite_logprob_tokens": [ + -4.147858142852783 + ], + "logprob_reward": -0.02285834215581417, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02285834215581417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20732258260250092, + 0.1945226937532425, + 0.019570428878068924, + 0.06556638330221176, + 0.0690496563911438, + 0.03594254329800606, + 0.046262770891189575, + 0.14690859615802765, + 0.09375649690628052, + 0.12109788507223129 + ], + "credit_map": [ + 0.022786596789956093, + 0.021387765184044838, + 0.03291964903473854, + 0.018113696947693825, + 0.01793551817536354, + 0.03457992523908615, + 0.029415911063551903, + 0.030183514580130577, + 0.019290052354335785, + 0.018782632425427437, + 0.028185613453388214, + 0.02197377197444439, + 0.012191868387162685, + 0.02725265920162201, + 0.03160031884908676, + 0.03524576500058174, + 0.0021711261942982674, + 0.0020311966072767973, + 0.003094386076554656, + 0.0021637368481606245, + 0.0010747527703642845, + 0.0024985172785818577, + 0.0033041907008737326, + 0.0032325200736522675, + 0.008548539131879807, + 0.006330673582851887, + 0.011359401047229767, + 0.012158725410699844, + 0.004139598459005356, + 0.00728203821927309, + 0.008086579851806164, + 0.007660821080207825, + 0.007825737819075584, + 0.0063797044567763805, + 0.01039276272058487, + 0.012959013693034649, + 0.004688628017902374, + 0.00809428095817566, + 0.008932395838201046, + 0.009777124039828777, + 0.005001232493668795, + 0.002784052398055792, + 0.007091302424669266, + 0.007353826425969601, + 0.001692474354058504, + 0.003918678965419531, + 0.0043575214222073555, + 0.0037434508558362722, + 0.0067182243801653385, + 0.004401599522680044, + 0.008641153573989868, + 0.009510710835456848, + 0.003002230543643236, + 0.003893825924023986, + 0.00635693222284317, + 0.003738089231774211, + 0.0195719376206398, + 0.010689279064536095, + 0.02813519537448883, + 0.024175161495804787, + 0.012099637649953365, + 0.014129631221294403, + 0.022967029362916946, + 0.015140705741941929, + 0.008230403065681458, + 0.011324271559715271, + 0.013938301242887974, + 0.015699047595262527, + 0.008035887964069843, + 0.01009439304471016, + 0.017123525962233543, + 0.009310654364526272, + 0.013404535129666328, + 0.022121774032711983, + 0.01586286723613739, + 0.02087966911494732, + 0.011312888003885746, + 0.007978837005794048, + 0.018257958814501762, + 0.011279347352683544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.02285834215581417 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005278697237372398 + ], + "opposite_logprob_tokens": [ + -5.255278587341309 + ], + "logprob_reward": -0.005278697237372398, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005278697237372398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07717636972665787, + 0.25550156831741333, + 0.06242753565311432, + 0.05468417704105377, + 0.024590758606791496, + 0.07238121330738068, + 0.041753124445676804, + 0.17893977463245392, + 0.12163877487182617, + 0.11090679466724396 + ], + "credit_map": [ + 0.011431542225182056, + 0.007845685817301273, + 0.01353593822568655, + 0.006377010606229305, + 0.004437878727912903, + 0.013083000667393208, + 0.00991481076925993, + 0.010550500825047493, + 0.02388983592391014, + 0.02166595496237278, + 0.03972119465470314, + 0.02667998895049095, + 0.017308780923485756, + 0.0424598790705204, + 0.035447437316179276, + 0.048328470438718796, + 0.006254616193473339, + 0.00503123365342617, + 0.010515058413147926, + 0.009214307181537151, + 0.004207677207887173, + 0.012369283474981785, + 0.008684193715453148, + 0.006151165347546339, + 0.005579672753810883, + 0.005519159138202667, + 0.008884681388735771, + 0.00793454609811306, + 0.0032298278529196978, + 0.007293562404811382, + 0.00920784194022417, + 0.0070348819717764854, + 0.002972702030092478, + 0.0020154742524027824, + 0.004079082049429417, + 0.004003788810223341, + 0.0016509939450770617, + 0.002840147353708744, + 0.003357636509463191, + 0.0036709352862089872, + 0.00782033521682024, + 0.006112414877861738, + 0.012329415418207645, + 0.013727913610637188, + 0.0050820824690163136, + 0.008600716479122639, + 0.008491740562021732, + 0.010216599330306053, + 0.006431900430470705, + 0.0031615421175956726, + 0.008106091059744358, + 0.0068411678075790405, + 0.003981863614171743, + 0.004267312120646238, + 0.005248097702860832, + 0.0037151528522372246, + 0.018729841336607933, + 0.014948457479476929, + 0.026658395305275917, + 0.03580917790532112, + 0.011015789583325386, + 0.017202317714691162, + 0.03479699417948723, + 0.019778799265623093, + 0.015770547091960907, + 0.014991351403295994, + 0.021675607189536095, + 0.02403268963098526, + 0.007520203944295645, + 0.00896901823580265, + 0.01622997224330902, + 0.012449372559785843, + 0.009892265312373638, + 0.01757311075925827, + 0.01465220469981432, + 0.014699527993798256, + 0.013628357090055943, + 0.00874252524226904, + 0.01982414722442627, + 0.011894658207893372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -0.005278697237372398 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983187697827816 + ], + "opposite_logprob_tokens": [ + -3.529831886291504 + ], + "logprob_reward": -0.02983187697827816, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983187697827816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19434045255184174, + 0.1537448912858963, + 0.3498378396034241, + 0.06364157050848007, + 0.05178939551115036, + 0.036682236939668655, + 0.02425806038081646, + 0.02764669805765152, + 0.0321260504424572, + 0.0659327358007431 + ], + "credit_map": [ + 0.021454017609357834, + 0.021941520273685455, + 0.03007754310965538, + 0.024963321164250374, + 0.01372440718114376, + 0.028787437826395035, + 0.0324370302259922, + 0.020955193787813187, + 0.015652626752853394, + 0.011701568029820919, + 0.02479509823024273, + 0.014714810065925121, + 0.009266006760299206, + 0.02459683082997799, + 0.019388433545827866, + 0.03362952917814255, + 0.036454398185014725, + 0.03378843516111374, + 0.050011973828077316, + 0.05111073702573776, + 0.028459839522838593, + 0.044320348650217056, + 0.061260879039764404, + 0.044431257992982864, + 0.006619278807193041, + 0.006357415113598108, + 0.010446342639625072, + 0.010441050864756107, + 0.003572981571778655, + 0.007626223377883434, + 0.009722545742988586, + 0.008855731226503849, + 0.00664726085960865, + 0.0077008819207549095, + 0.007079598028212786, + 0.007836319506168365, + 0.005106394179165363, + 0.005832540802657604, + 0.0060969009064137936, + 0.0054895030334591866, + 0.0036439138930290937, + 0.0028141813818365335, + 0.0054584164172410965, + 0.005745397415012121, + 0.002807024633511901, + 0.004736227449029684, + 0.005058364477008581, + 0.006418714765459299, + 0.002753599314019084, + 0.0021445383317768574, + 0.0037493263371288776, + 0.0043973722495138645, + 0.002368117682635784, + 0.003107387339696288, + 0.003302630502730608, + 0.0024350902531296015, + 0.0032962418626993895, + 0.0022491507697850466, + 0.00423998711630702, + 0.004884714260697365, + 0.002060161903500557, + 0.0028703981079161167, + 0.004759042523801327, + 0.0032870054710656404, + 0.003483438864350319, + 0.004406733904033899, + 0.004067983012646437, + 0.004183061886578798, + 0.0031424316111952066, + 0.0031557148322463036, + 0.006053989287465811, + 0.003632699605077505, + 0.005774604622274637, + 0.0076605104841291904, + 0.010273866355419159, + 0.006944866850972176, + 0.007357632741332054, + 0.006329525262117386, + 0.011302073486149311, + 0.010289667174220085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.02983187697827816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981729805469513 + ], + "opposite_logprob_tokens": [ + -3.5298173427581787 + ], + "logprob_reward": -0.02981729805469513, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981729805469513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14853131771087646, + 0.09840283542871475, + 0.040898941457271576, + 0.03396633639931679, + 0.03009103238582611, + 0.04450501129031181, + 0.029589086771011353, + 0.12161858379840851, + 0.19252634048461914, + 0.2598705291748047 + ], + "credit_map": [ + 0.018055744469165802, + 0.015803523361682892, + 0.02545788325369358, + 0.018732963129878044, + 0.009299333207309246, + 0.025731872767210007, + 0.018767721951007843, + 0.01668229140341282, + 0.012826623395085335, + 0.00762717192992568, + 0.014927001670002937, + 0.007475301623344421, + 0.004364619497209787, + 0.011268887668848038, + 0.014533930458128452, + 0.025379294529557228, + 0.004765847232192755, + 0.0038898279890418053, + 0.00596190569922328, + 0.006750572007149458, + 0.002962477272376418, + 0.003252544905990362, + 0.008281542919576168, + 0.005034222267568111, + 0.004477459937334061, + 0.0031162865925580263, + 0.00547535065561533, + 0.004117656499147415, + 0.0017566595925018191, + 0.004069115966558456, + 0.005785365588963032, + 0.00516844354569912, + 0.00364707806147635, + 0.003060998162254691, + 0.005017254501581192, + 0.004938384983688593, + 0.001809716341085732, + 0.003038056194782257, + 0.004030388314276934, + 0.004549157340079546, + 0.006023583933711052, + 0.0031750372145324945, + 0.009238163940608501, + 0.008764958940446377, + 0.002248425269499421, + 0.006247639190405607, + 0.005128978285938501, + 0.003678223816677928, + 0.004948793910443783, + 0.002659009536728263, + 0.005415426101535559, + 0.006607663352042437, + 0.0014278320595622063, + 0.002869281219318509, + 0.0033510089851915836, + 0.002310069277882576, + 0.011565213091671467, + 0.00985704455524683, + 0.01911258138716221, + 0.015390094369649887, + 0.012221971526741982, + 0.011477966792881489, + 0.029339617118239403, + 0.012654089368879795, + 0.019937364384531975, + 0.03345983102917671, + 0.028331680223345757, + 0.029954254627227783, + 0.013363623060286045, + 0.020034629851579666, + 0.027885792776942253, + 0.019559159874916077, + 0.024362927302718163, + 0.052801404148340225, + 0.03279517590999603, + 0.027295857667922974, + 0.03811049088835716, + 0.020893946290016174, + 0.03794512525200844, + 0.025665605440735817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.02981729805469513 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene framed as an extremely close, still photograph held flush against a dark, polished oak table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an off-white paper visible, crumpled into sharp creases and soft wrinkles, balancing precariously on one edge?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a single warm light source visible on the right side of the frame casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the paper remain perfectly still for a moment before an almost imperceptible tremor causes it to collapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paper fall flat onto the wooden surface in a soft, silent motion, reducing to a two-dimensional crumpled sheet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002824010793119669 + ], + "opposite_logprob_tokens": [ + -5.877823829650879 + ], + "logprob_reward": -0.002824010793119669, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002824010793119669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11418195813894272, + 0.041752126067876816, + 0.02572670578956604, + 0.024103451520204544, + 0.023497020825743675, + 0.0547235831618309, + 0.07870039343833923, + 0.12701354920864105, + 0.2269831746816635, + 0.28331804275512695 + ], + "credit_map": [ + 0.01251148246228695, + 0.011776421219110489, + 0.01755920611321926, + 0.011054730974137783, + 0.0092738326638937, + 0.023400789126753807, + 0.014607526361942291, + 0.01399796362966299, + 0.006471150554716587, + 0.0038925071712583303, + 0.006926888134330511, + 0.003901722142472863, + 0.002248888835310936, + 0.007355337496846914, + 0.006122170947492123, + 0.004833463113754988, + 0.003376295091584325, + 0.0020057058427482843, + 0.004268045071512461, + 0.003557942109182477, + 0.001456624362617731, + 0.0026352007407695055, + 0.00482055963948369, + 0.003606333862990141, + 0.003595570335164666, + 0.0023688864894211292, + 0.004141949117183685, + 0.003164642257615924, + 0.0015707265119999647, + 0.0032931494060903788, + 0.003032284788787365, + 0.0029362435452640057, + 0.0034704403951764107, + 0.002867260714992881, + 0.00485058780759573, + 0.003382559400051832, + 0.0018228142289444804, + 0.002912633353844285, + 0.0026543792337179184, + 0.0015363484853878617, + 0.0066144950687885284, + 0.005376308225095272, + 0.00998599547892809, + 0.011172044090926647, + 0.003712515579536557, + 0.006033408921211958, + 0.00677616149187088, + 0.005052654538303614, + 0.009355246089398861, + 0.00664094602689147, + 0.012467381544411182, + 0.01518149022012949, + 0.006166728213429451, + 0.01128357369452715, + 0.009925984777510166, + 0.007679041009396315, + 0.014301416464149952, + 0.012018018402159214, + 0.023486938327550888, + 0.02638658881187439, + 0.007813360542058945, + 0.01119684986770153, + 0.01874524913728237, + 0.013065135106444359, + 0.023859938606619835, + 0.040599048137664795, + 0.035000573843717575, + 0.03355109319090843, + 0.01400911994278431, + 0.027607686817646027, + 0.030373014509677887, + 0.021982694044709206, + 0.023665519431233406, + 0.04376370087265968, + 0.03959058225154877, + 0.04595724493265152, + 0.031705014407634735, + 0.024274181574583054, + 0.03983130678534508, + 0.034530479460954666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.002824010793119669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_7/index_tensor([3508, 3508])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_7/index_tensor([3508, 3508])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5b2cf0701959c3165da745d52814df7e2dd3cdef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_7/index_tensor([3508, 3508])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic drone shot slowly zooms forward from high above, tilting slightly upward to reveal a vast vista of a deep canyon, stained golden red by the setting afternoon sun. In the canyon's center, a mountainous skeleton of an ancient creature sits hunched in a posture of eternal contemplation. Its massive, rock-like bones are a weathered off-white, covered in deep cracks and mottled with dark moss, its massive spine curved in a mournful arc. The steep, crimson rock walls cast long, deep shadows, partially shrouding the skeleton in a mysterious darkness, while sunlight sharply outlines the contours of its skull and shoulder blades. A strong wind blew steadily and continuously, sweeping up the dry yellow sand on the ground and forming clearly visible sand flows. These sand flows seemed to have life, whistling and flowing unhindered through the abyss-like empty eye sockets of the skull, and steadily pouring out from the wide gaps between the ribs of its magnificent cathedral dome, forming a misty and continuously flowing sand fog on the leeward side of the skeleton. The whole picture showed movement in an extreme silence, full of solemn, desolate and extremely beautiful atmosphere, conveying an epic sense of loneliness spanning the ages.", + "seed": 15128, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_7/index_3508_seed_15128_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.939701995998621, + "vlm_logprob": -0.939701995998621, + "vlm_yes_no_margin": 0.825000011920929, + "vlm_token_credit": -0.939701995998621, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -3.168850898742676 + ], + "opposite_logprob_tokens": [ + -0.0438508503139019 + ], + "logprob_reward": -3.168850898742676, + "yes_no_margin_reward": -3.125, + "token_credit_reward": -3.168850898742676, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18039998412132263, + 0.09061184525489807, + 0.1512235701084137, + 0.1308746635913849, + 0.08151751011610031, + 0.16420048475265503, + 0.019227320328354836, + 0.020528078079223633, + 0.05800818279385567, + 0.10340835899114609 + ], + "credit_map": [ + 0.03869063779711723, + 0.014609061181545258, + 0.021384891122579575, + 0.01908673346042633, + 0.011837968602776527, + 0.01975618302822113, + 0.01005820743739605, + 0.04497629776597023, + 0.013847964815795422, + 0.005239078309386969, + 0.020839165896177292, + 0.009332066401839256, + 0.0046053663827478886, + 0.015038828365504742, + 0.00976121611893177, + 0.011948156170547009, + 0.027791516855359077, + 0.013749483041465282, + 0.012094078585505486, + 0.017650650814175606, + 0.01121768169105053, + 0.02341916784644127, + 0.01532670296728611, + 0.02997428923845291, + 0.024645157158374786, + 0.009733659215271473, + 0.019393956288695335, + 0.011004929430782795, + 0.009586178697645664, + 0.03249902278184891, + 0.01293446309864521, + 0.01107729785144329, + 0.009139100089669228, + 0.01060501765459776, + 0.013417856767773628, + 0.00794935505837202, + 0.005530116613954306, + 0.016647491604089737, + 0.008926020935177803, + 0.009302549064159393, + 0.020399952307343483, + 0.011520655825734138, + 0.018820900470018387, + 0.01614232175052166, + 0.010593341663479805, + 0.043388474732637405, + 0.024105357006192207, + 0.019229480996727943, + 0.002135374117642641, + 0.00259860185906291, + 0.0020016010385006666, + 0.0024858731776475906, + 0.0023761189077049494, + 0.0034251557663083076, + 0.002510109217837453, + 0.0016944862436503172, + 0.002214709995314479, + 0.0025276942178606987, + 0.0022405972704291344, + 0.002702812198549509, + 0.0014919164823368192, + 0.004151043947786093, + 0.0030898803379386663, + 0.0021094244439154863, + 0.006301229353994131, + 0.003534352872520685, + 0.006698885001242161, + 0.003593619680032134, + 0.004105168394744396, + 0.01957280933856964, + 0.007586117368191481, + 0.006615997292101383, + 0.009861585684120655, + 0.012667140923440456, + 0.012139786034822464, + 0.012015539221465588, + 0.010547561571002007, + 0.021636711433529854, + 0.013996027410030365, + 0.0105440029874444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -3.168850898742676 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3136730194091797 + ], + "opposite_logprob_tokens": [ + -0.31367307901382446 + ], + "logprob_reward": -1.3136730194091797, + "yes_no_margin_reward": -0.9999999403953552, + "token_credit_reward": -1.3136730194091797, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2324080914258957, + 0.08312630653381348, + 0.3369210660457611, + 0.09265469014644623, + 0.050124768167734146, + 0.0555223785340786, + 0.02852519601583481, + 0.02141190506517887, + 0.06223658099770546, + 0.037069011479616165 + ], + "credit_map": [ + 0.04227341338992119, + 0.013227179646492004, + 0.02796347439289093, + 0.020200731232762337, + 0.010741977021098137, + 0.029170487076044083, + 0.013556181453168392, + 0.07527464628219604, + 0.018608009442687035, + 0.006660187151283026, + 0.008316065184772015, + 0.008650883100926876, + 0.005726768635213375, + 0.013313909992575645, + 0.007805750705301762, + 0.01404473278671503, + 0.06290854513645172, + 0.021673128008842468, + 0.0294629018753767, + 0.02720906026661396, + 0.020608048886060715, + 0.08010069280862808, + 0.027735363692045212, + 0.06722333282232285, + 0.014125092886388302, + 0.010235778987407684, + 0.013253651559352875, + 0.011121192015707493, + 0.006284752860665321, + 0.017273198813199997, + 0.010007991455495358, + 0.010353033430874348, + 0.0065114544704556465, + 0.00520940450951457, + 0.005663675721734762, + 0.006156783550977707, + 0.004463297314941883, + 0.011233710683882236, + 0.0056227208115160465, + 0.005263722501695156, + 0.0068914564326405525, + 0.004583651199936867, + 0.006526454351842403, + 0.005811802577227354, + 0.0039767432026565075, + 0.01556041557341814, + 0.0060903048142790794, + 0.006081549450755119, + 0.0037707837764173746, + 0.0025714831426739693, + 0.0029551631305366755, + 0.002328452654182911, + 0.0016290502389892936, + 0.008424720726907253, + 0.004382723942399025, + 0.002462818520143628, + 0.00309215416200459, + 0.00235160649754107, + 0.0021934930700808764, + 0.00190857017878443, + 0.0019199930829927325, + 0.0054877386428415775, + 0.002343402709811926, + 0.002114948583766818, + 0.006220177281647921, + 0.005129744298756123, + 0.005791081581264734, + 0.005873174872249365, + 0.005216753110289574, + 0.017599860206246376, + 0.00972941517829895, + 0.006676381453871727, + 0.004011852201074362, + 0.004392317030578852, + 0.005365213379263878, + 0.005359536502510309, + 0.003093270119279623, + 0.006485230289399624, + 0.003982345573604107, + 0.004379246383905411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -1.3136730194091797 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033795833587646484 + ], + "opposite_logprob_tokens": [ + -3.4087958335876465 + ], + "logprob_reward": -0.033795833587646484, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033795833587646484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14133693277835846, + 0.1872752159833908, + 0.2346709668636322, + 0.044194646179676056, + 0.033802613615989685, + 0.07416048645973206, + 0.011832201853394508, + 0.04651472717523575, + 0.04574631527066231, + 0.18046587705612183 + ], + "credit_map": [ + 0.016570284962654114, + 0.012809843756258488, + 0.026534292846918106, + 0.016676345840096474, + 0.01310428511351347, + 0.021762698888778687, + 0.01253924984484911, + 0.021339943632483482, + 0.02527112513780594, + 0.008922839537262917, + 0.062477316707372665, + 0.01645537279546261, + 0.008413908071815968, + 0.03340315446257591, + 0.014441064558923244, + 0.01789044961333275, + 0.038663726300001144, + 0.020231714472174644, + 0.020897163078188896, + 0.025949331000447273, + 0.019623668864369392, + 0.05827794224023819, + 0.031045813113451004, + 0.019981607794761658, + 0.006871597841382027, + 0.005303387995809317, + 0.005676005966961384, + 0.004817441571503878, + 0.003300022566691041, + 0.007955160923302174, + 0.00587947852909565, + 0.004391551483422518, + 0.004655842669308186, + 0.003626333549618721, + 0.0033625117503106594, + 0.003613349050283432, + 0.002616472775116563, + 0.007994182407855988, + 0.0033251147251576185, + 0.004608808550983667, + 0.008484387770295143, + 0.005061566364020109, + 0.007820880971848965, + 0.007151423953473568, + 0.0052666994743049145, + 0.019880011677742004, + 0.012083398178219795, + 0.008412121795117855, + 0.0014810280408710241, + 0.0013960490468889475, + 0.0011574174277484417, + 0.0017690346576273441, + 0.0013999150833114982, + 0.0021511628292500973, + 0.0013268928742036223, + 0.001150702708400786, + 0.0057183485478162766, + 0.005648224148899317, + 0.004467136692255735, + 0.004853371996432543, + 0.00530234957113862, + 0.008105156943202019, + 0.007409799378365278, + 0.005010346416383982, + 0.004778062924742699, + 0.0043891556560993195, + 0.0053703393787145615, + 0.0034610514994710684, + 0.004765711259096861, + 0.013073955662548542, + 0.004822758957743645, + 0.005085285287350416, + 0.01838887296617031, + 0.023423058912158012, + 0.02394229918718338, + 0.022583400830626488, + 0.018583977594971657, + 0.03226930648088455, + 0.023226017132401466, + 0.018048960715532303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.033795833587646484 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05502470210194588 + ], + "opposite_logprob_tokens": [ + -2.9300246238708496 + ], + "logprob_reward": -0.05502470210194588, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05502470210194588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3080551028251648, + 0.08114548772573471, + 0.09678301960229874, + 0.07310307770967484, + 0.10551021993160248, + 0.044221166521310806, + 0.025427935644984245, + 0.05594882369041443, + 0.038740236312150955, + 0.17106488347053528 + ], + "credit_map": [ + 0.06957954168319702, + 0.013798925094306469, + 0.03490700200200081, + 0.02698383666574955, + 0.007725760340690613, + 0.030499661341309547, + 0.012160892598330975, + 0.11239948868751526, + 0.013608275912702084, + 0.006404700223356485, + 0.012104887515306473, + 0.007259156089276075, + 0.006968231871724129, + 0.01671617291867733, + 0.006619682535529137, + 0.011464380659162998, + 0.01665372960269451, + 0.009538351558148861, + 0.008916765451431274, + 0.01327646803110838, + 0.006950004026293755, + 0.012988434173166752, + 0.011402868665754795, + 0.017056403681635857, + 0.012218708172440529, + 0.004227683879435062, + 0.013749725185334682, + 0.007417322602123022, + 0.0034549348056316376, + 0.017205018550157547, + 0.005734774749726057, + 0.009094915352761745, + 0.017679747194051743, + 0.0078116524964571, + 0.01366902980953455, + 0.007299643009901047, + 0.006884029135107994, + 0.03387738764286041, + 0.007330219727009535, + 0.010958511382341385, + 0.0050755711272358894, + 0.0031461999751627445, + 0.005377619527280331, + 0.004854162223637104, + 0.0030189284589141607, + 0.010842923074960709, + 0.006699512246996164, + 0.005206251982599497, + 0.0031478586606681347, + 0.0026283590123057365, + 0.0027402390260249376, + 0.0024777615908533335, + 0.002303841058164835, + 0.00693910289555788, + 0.002721228636801243, + 0.002469545230269432, + 0.007482732646167278, + 0.0068503934890031815, + 0.005134279374033213, + 0.0063495151698589325, + 0.006620423868298531, + 0.010390409268438816, + 0.007172033190727234, + 0.005949035752564669, + 0.003632789710536599, + 0.00517446780577302, + 0.0049807364121079445, + 0.003841178957372904, + 0.004582502413541079, + 0.007373034954071045, + 0.004902057349681854, + 0.00425346614792943, + 0.02695428393781185, + 0.016954118385910988, + 0.018481187522411346, + 0.01342182606458664, + 0.011169358156621456, + 0.0410747230052948, + 0.022683674469590187, + 0.020325714722275734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.05502470210194588 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1271655261516571 + ], + "opposite_logprob_tokens": [ + -2.1271655559539795 + ], + "logprob_reward": -0.1271655261516571, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1271655261516571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15479011833667755, + 0.2284436821937561, + 0.22679810225963593, + 0.07603864371776581, + 0.0713832899928093, + 0.021435357630252838, + 0.04358622431755066, + 0.04411246255040169, + 0.04047172889113426, + 0.09294042736291885 + ], + "credit_map": [ + 0.023390641435980797, + 0.014274056069552898, + 0.016939591616392136, + 0.017760304734110832, + 0.013684380799531937, + 0.016613619402050972, + 0.013983987271785736, + 0.038143523037433624, + 0.03676348552107811, + 0.021985521540045738, + 0.031533095985651016, + 0.021843498572707176, + 0.014627155847847462, + 0.04050309956073761, + 0.021452367305755615, + 0.03973542898893356, + 0.03861253336071968, + 0.017047327011823654, + 0.02188274636864662, + 0.02207696996629238, + 0.015710774809122086, + 0.06001776084303856, + 0.01950688473880291, + 0.031943097710609436, + 0.01060763094574213, + 0.008029544726014137, + 0.010755134746432304, + 0.010074961930513382, + 0.004541547503322363, + 0.017052985727787018, + 0.00822181161493063, + 0.006755022332072258, + 0.009371096268296242, + 0.009725244715809822, + 0.009346149861812592, + 0.007935580797493458, + 0.005251712165772915, + 0.017989451065659523, + 0.005743874236941338, + 0.006020172964781523, + 0.0023361211642622948, + 0.0022387756034731865, + 0.002556151943281293, + 0.002919950522482395, + 0.001531208399683237, + 0.004644531290978193, + 0.0030886102467775345, + 0.0021200058981776237, + 0.005877628922462463, + 0.0037420587614178658, + 0.0036394521594047546, + 0.004868020303547382, + 0.0028764375019818544, + 0.014386199414730072, + 0.004302718676626682, + 0.0038937025237828493, + 0.005656793247908354, + 0.006073339376598597, + 0.0027857599779963493, + 0.005410319194197655, + 0.004557412583380938, + 0.009033242240548134, + 0.006455766037106514, + 0.004139828495681286, + 0.004338499624282122, + 0.003745130030438304, + 0.003467490430921316, + 0.003931066486984491, + 0.003083480754867196, + 0.01305166445672512, + 0.004258586559444666, + 0.004595810081809759, + 0.0105286231264472, + 0.01712885871529579, + 0.011635404080152512, + 0.01049865037202835, + 0.011096148751676083, + 0.012178155593574047, + 0.009251737035810947, + 0.010622833855450153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.1271655261516571 + } + ], + "vlm_decoded_outputs": [ + "No", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic drone shot slowly zooms forward from high above, tilting slightly upward to reveal a vast vista of a deep canyon, stained golden red by the setting afternoon sun. In the canyon's center, a mountainous skeleton of an ancient creature sits hunched in a posture of eternal contemplation. Its massive, rock-like bones are a weathered off-white, covered in deep cracks and mottled with dark moss, its massive spine curved in a mournful arc. The steep, crimson rock walls cast long, deep shadows, partially shrouding the skeleton in a mysterious darkness, while sunlight sharply outlines the contours of its skull and shoulder blades. A strong wind blew steadily and continuously, sweeping up the dry yellow sand on the ground and forming clearly visible sand flows. These sand flows seemed to have life, whistling and flowing unhindered through the abyss-like empty eye sockets of the skull, and steadily pouring out from the wide gaps between the ribs of its magnificent cathedral dome, forming a misty and continuously flowing sand fog on the leeward side of the skeleton. The whole picture showed movement in an extreme silence, full of solemn, desolate and extremely beautiful atmosphere, conveying an epic sense of loneliness spanning the ages.", + "seed": 15129, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_7/index_3508_seed_15129_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.7235531575977803, + "vlm_logprob": -0.7235531575977803, + "vlm_yes_no_margin": 0.975, + "vlm_token_credit": -0.7235531575977803, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.4648759365081787 + ], + "opposite_logprob_tokens": [ + -0.08987598866224289 + ], + "logprob_reward": -2.4648759365081787, + "yes_no_margin_reward": -2.375, + "token_credit_reward": -2.4648759365081787, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10381003469228745, + 0.07693114131689072, + 0.06970684975385666, + 0.4312557578086853, + 0.07505081593990326, + 0.03761900216341019, + 0.05387146398425102, + 0.01974756456911564, + 0.07733066380023956, + 0.054676733911037445 + ], + "credit_map": [ + 0.011733971536159515, + 0.00849162694066763, + 0.009525499306619167, + 0.009318690747022629, + 0.007464142981916666, + 0.00544116273522377, + 0.02905471809208393, + 0.022780215367674828, + 0.008713479153811932, + 0.007038201671093702, + 0.007659596391022205, + 0.007101964671164751, + 0.006137312389910221, + 0.0074298265390098095, + 0.013991975225508213, + 0.018858786672353745, + 0.00864998809993267, + 0.00768633047118783, + 0.007246985565871, + 0.007790693547576666, + 0.00742614921182394, + 0.00697782589122653, + 0.011250189505517483, + 0.012678690254688263, + 0.01762847788631916, + 0.006021077744662762, + 0.02066163346171379, + 0.00868619792163372, + 0.009642046876251698, + 0.009777732193470001, + 0.18071633577346802, + 0.17812226712703705, + 0.00848125759512186, + 0.010665358044207096, + 0.017314281314611435, + 0.008010756224393845, + 0.007040756288915873, + 0.005061479285359383, + 0.009926936589181423, + 0.008549997583031654, + 0.0030500052962452173, + 0.005886994302272797, + 0.00590941496193409, + 0.00559624657034874, + 0.004285003989934921, + 0.0034154083114117384, + 0.004799299873411655, + 0.0046766288578510284, + 0.007013911847025156, + 0.0056948852725327015, + 0.009505764581263065, + 0.009493302553892136, + 0.003730860771611333, + 0.0046155634336173534, + 0.007937559857964516, + 0.005879617296159267, + 0.0023019893560558558, + 0.0023593679070472717, + 0.003034126479178667, + 0.0034394231624901295, + 0.0020600263960659504, + 0.0020753287244588137, + 0.0019184559350833297, + 0.002558846492320299, + 0.0068297432735562325, + 0.007525534834712744, + 0.013145972043275833, + 0.01918911002576351, + 0.006277511827647686, + 0.007144507486373186, + 0.006121563259512186, + 0.01109672337770462, + 0.007001352962106466, + 0.0066073620691895485, + 0.010124359279870987, + 0.007403156254440546, + 0.004776959773153067, + 0.006175411865115166, + 0.0063368394039571285, + 0.006251287180930376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -2.4648759365081787 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8985159397125244 + ], + "opposite_logprob_tokens": [ + -0.5235159397125244 + ], + "logprob_reward": -0.8985159397125244, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8985159397125244, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08823174238204956, + 0.13435900211334229, + 0.056675903499126434, + 0.16544774174690247, + 0.13041876256465912, + 0.04217572137713432, + 0.06416084617376328, + 0.05078744888305664, + 0.15263399481773376, + 0.11510880291461945 + ], + "credit_map": [ + 0.012109470553696156, + 0.009931228123605251, + 0.009570625610649586, + 0.012561162933707237, + 0.014467290602624416, + 0.006520113442093134, + 0.013686269521713257, + 0.009385587647557259, + 0.010724682360887527, + 0.01931428723037243, + 0.019503291696310043, + 0.02272089757025242, + 0.015793412923812866, + 0.01130311843007803, + 0.017478495836257935, + 0.01752082072198391, + 0.007334067486226559, + 0.007008790038526058, + 0.005406225100159645, + 0.005678605288267136, + 0.00929485447704792, + 0.005515333730727434, + 0.006914543453603983, + 0.009523482993245125, + 0.007374972105026245, + 0.0033394037745893, + 0.00923356506973505, + 0.003886365331709385, + 0.004126602318137884, + 0.004628003109246492, + 0.06737405806779861, + 0.06548476964235306, + 0.015701092779636383, + 0.01355455070734024, + 0.02232942543923855, + 0.01291830837726593, + 0.013505333103239536, + 0.012975194491446018, + 0.020611794665455818, + 0.01882307045161724, + 0.003230396658182144, + 0.007292109541594982, + 0.0058437311090528965, + 0.0063452343456447124, + 0.004568181466311216, + 0.0042643435299396515, + 0.00574957299977541, + 0.0048821549862623215, + 0.006870751269161701, + 0.0054100495763123035, + 0.011568439193069935, + 0.011452021077275276, + 0.004771706648170948, + 0.006033648736774921, + 0.009390715509653091, + 0.008663512766361237, + 0.005530182737857103, + 0.004334901459515095, + 0.010706961154937744, + 0.011058573611080647, + 0.0025996461044996977, + 0.003601046511903405, + 0.006151982117444277, + 0.006804157048463821, + 0.014650998637080193, + 0.014943393878638744, + 0.023066936060786247, + 0.022309282794594765, + 0.01830081082880497, + 0.024599870666861534, + 0.012494786642491817, + 0.022267915308475494, + 0.015066429041326046, + 0.016476716846227646, + 0.020969154313206673, + 0.017789898440241814, + 0.009023504331707954, + 0.00891675055027008, + 0.00817785318940878, + 0.01868850365281105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.8985159397125244 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038150887936353683 + ], + "opposite_logprob_tokens": [ + -3.2881507873535156 + ], + "logprob_reward": -0.038150887936353683, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038150887936353683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08543435484170914, + 0.05457692965865135, + 0.09865052998065948, + 0.4689962565898895, + 0.06528487801551819, + 0.032656215131282806, + 0.039977751672267914, + 0.046503279358148575, + 0.06384101510047913, + 0.04407874122262001 + ], + "credit_map": [ + 0.014697962440550327, + 0.008618555031716824, + 0.011996612884104252, + 0.012911940924823284, + 0.006579605396836996, + 0.006666939239948988, + 0.011542762629687786, + 0.01241997815668583, + 0.004197925329208374, + 0.006176885217428207, + 0.006631493102759123, + 0.007431495003402233, + 0.004351553972810507, + 0.00408177962526679, + 0.011231868527829647, + 0.010473928414285183, + 0.007791948039084673, + 0.010175304487347603, + 0.01202385127544403, + 0.014481846243143082, + 0.0073440950363874435, + 0.006546104326844215, + 0.020866334438323975, + 0.01942105032503605, + 0.019185977056622505, + 0.008739596232771873, + 0.02188895270228386, + 0.013516871258616447, + 0.012529975734651089, + 0.01030754018574953, + 0.19169862568378448, + 0.1911287158727646, + 0.008359991945326328, + 0.007306597661226988, + 0.014965720474720001, + 0.006269770208746195, + 0.006179352290928364, + 0.005002287682145834, + 0.008612646721303463, + 0.008588515222072601, + 0.0025765225291252136, + 0.005602026358246803, + 0.0054711587727069855, + 0.0038547110743820667, + 0.004196555819362402, + 0.002692791633307934, + 0.004245921038091183, + 0.004016529303044081, + 0.004579990636557341, + 0.00462855352088809, + 0.005886055529117584, + 0.0072466423735022545, + 0.0038474833127111197, + 0.0034204546827822924, + 0.0056159161031246185, + 0.004752654582262039, + 0.005163457710295916, + 0.006038882303982973, + 0.00798381119966507, + 0.008463364094495773, + 0.003933372441679239, + 0.004103026352822781, + 0.004977294709533453, + 0.005840074736624956, + 0.006722288206219673, + 0.006323055364191532, + 0.01044158823788166, + 0.013144322670996189, + 0.00505916029214859, + 0.008162787184119225, + 0.005275012459605932, + 0.008712797425687313, + 0.005400977097451687, + 0.004926119931042194, + 0.007616991642862558, + 0.00477999122813344, + 0.0054096351377666, + 0.005342667922377586, + 0.005559297278523445, + 0.005043062847107649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.038150887936353683 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909224718809128 + ], + "opposite_logprob_tokens": [ + -2.464092254638672 + ], + "logprob_reward": -0.08909224718809128, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909224718809128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13781721889972687, + 0.06699790060520172, + 0.03525163233280182, + 0.4231826066970825, + 0.14491385221481323, + 0.01691974699497223, + 0.07206213474273682, + 0.05218847841024399, + 0.031939659267663956, + 0.01872672513127327 + ], + "credit_map": [ + 0.012578424997627735, + 0.009683135896921158, + 0.012645035982131958, + 0.01184658519923687, + 0.014446117915213108, + 0.007137381937354803, + 0.04184727743268013, + 0.027633264660835266, + 0.0062377844005823135, + 0.00469024945050478, + 0.009336592629551888, + 0.007836529985070229, + 0.004340717568993568, + 0.006798097863793373, + 0.013621507212519646, + 0.014136423356831074, + 0.0021421995479613543, + 0.0028112241998314857, + 0.004355871118605137, + 0.0024391692131757736, + 0.003533876035362482, + 0.0021833342034369707, + 0.008398057892918587, + 0.009387901984155178, + 0.017530789598822594, + 0.0080342348664999, + 0.019948231056332588, + 0.01083358097821474, + 0.011133205145597458, + 0.011093469336628914, + 0.17283260822296143, + 0.17177650332450867, + 0.01836935617029667, + 0.021346649155020714, + 0.028976740315556526, + 0.01730799674987793, + 0.011504711583256721, + 0.012319241650402546, + 0.018129907548427582, + 0.01695925183594227, + 0.00169229821767658, + 0.0019091935828328133, + 0.0025532173458486795, + 0.0018563830526545644, + 0.0027036804240196943, + 0.0015410592313855886, + 0.002573537640273571, + 0.002090378664433956, + 0.009260931052267551, + 0.007400249596685171, + 0.009591766633093357, + 0.015614383853971958, + 0.00614300649613142, + 0.006572275422513485, + 0.009403125382959843, + 0.008076400496065617, + 0.0061251455917954445, + 0.005614357069134712, + 0.008226858451962471, + 0.00930191483348608, + 0.00511338422074914, + 0.00501682935282588, + 0.006557268090546131, + 0.006232724990695715, + 0.003929551225155592, + 0.003158728126436472, + 0.003733200253918767, + 0.0059051429852843285, + 0.0033219580072909594, + 0.004156036302447319, + 0.002555446233600378, + 0.005179595667868853, + 0.0022780373692512512, + 0.002037936821579933, + 0.003808623878285289, + 0.0028742472641170025, + 0.0015676276525482535, + 0.0020664797630161047, + 0.0018076952546834946, + 0.0022860784083604813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.08909224718809128 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12713077664375305 + ], + "opposite_logprob_tokens": [ + -2.1271307468414307 + ], + "logprob_reward": -0.12713077664375305, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12713077664375305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07224539667367935, + 0.0841245949268341, + 0.04742719605565071, + 0.26581814885139465, + 0.16083350777626038, + 0.06184076517820358, + 0.17891721427440643, + 0.047126948833465576, + 0.04969940334558487, + 0.031966835260391235 + ], + "credit_map": [ + 0.010268548503518105, + 0.005281350575387478, + 0.007819807156920433, + 0.008297594264149666, + 0.004403302446007729, + 0.0038569653406739235, + 0.016345813870429993, + 0.015972016379237175, + 0.008706042543053627, + 0.01168555486947298, + 0.012430750764906406, + 0.011178665794432163, + 0.010931993834674358, + 0.0075648631900548935, + 0.01153904851526022, + 0.010087675414979458, + 0.0037863075267523527, + 0.006201952695846558, + 0.006915336009114981, + 0.006595146376639605, + 0.007190907374024391, + 0.004528481047600508, + 0.005938318558037281, + 0.006270750891417265, + 0.010888054966926575, + 0.005523176863789558, + 0.014427910558879375, + 0.0058479285798966885, + 0.006045445334166288, + 0.006308649200946093, + 0.10979873687028885, + 0.10697824507951736, + 0.0162142813205719, + 0.020897161215543747, + 0.03988947346806526, + 0.01770717464387417, + 0.015407469123601913, + 0.010924260132014751, + 0.021105531603097916, + 0.018688151612877846, + 0.005320598836988211, + 0.010304184630513191, + 0.009444205090403557, + 0.008566701784729958, + 0.008299954235553741, + 0.004988397005945444, + 0.0074173142202198505, + 0.007499408908188343, + 0.02490433305501938, + 0.016766248270869255, + 0.028879165649414062, + 0.03644636645913124, + 0.011870105750858784, + 0.013854816555976868, + 0.0216217078268528, + 0.024574482813477516, + 0.00555776059627533, + 0.004634740296751261, + 0.006833484396338463, + 0.006143397185951471, + 0.005591718014329672, + 0.005167872179299593, + 0.0067453533411026, + 0.00645262049511075, + 0.004015333950519562, + 0.004653396084904671, + 0.006430317647755146, + 0.012645436450839043, + 0.004575714934617281, + 0.005165024194866419, + 0.004750940948724747, + 0.0074632428586483, + 0.003204531501978636, + 0.003436885541304946, + 0.0058775171637535095, + 0.004783965181559324, + 0.0035892154555767775, + 0.0029953571502119303, + 0.0034255259670317173, + 0.004653837997466326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.12713077664375305 + } + ], + "vlm_decoded_outputs": [ + "No", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic drone shot slowly zooms forward from high above, tilting slightly upward to reveal a vast vista of a deep canyon, stained golden red by the setting afternoon sun. In the canyon's center, a mountainous skeleton of an ancient creature sits hunched in a posture of eternal contemplation. Its massive, rock-like bones are a weathered off-white, covered in deep cracks and mottled with dark moss, its massive spine curved in a mournful arc. The steep, crimson rock walls cast long, deep shadows, partially shrouding the skeleton in a mysterious darkness, while sunlight sharply outlines the contours of its skull and shoulder blades. A strong wind blew steadily and continuously, sweeping up the dry yellow sand on the ground and forming clearly visible sand flows. These sand flows seemed to have life, whistling and flowing unhindered through the abyss-like empty eye sockets of the skull, and steadily pouring out from the wide gaps between the ribs of its magnificent cathedral dome, forming a misty and continuously flowing sand fog on the leeward side of the skeleton. The whole picture showed movement in an extreme silence, full of solemn, desolate and extremely beautiful atmosphere, conveying an epic sense of loneliness spanning the ages.", + "seed": 15130, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_7/index_3508_seed_15130_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.49145060796290635, + "vlm_logprob": -0.49145060796290635, + "vlm_yes_no_margin": 1.7500000178813935, + "vlm_token_credit": -0.49145060796290635, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.315321922302246 + ], + "opposite_logprob_tokens": [ + -0.31532198190689087 + ], + "logprob_reward": -1.315321922302246, + "yes_no_margin_reward": -0.9999999403953552, + "token_credit_reward": -1.315321922302246, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19653446972370148, + 0.11540593951940536, + 0.19823166728019714, + 0.015588905662298203, + 0.2894008457660675, + 0.01843806356191635, + 0.03516471013426781, + 0.051985956728458405, + 0.053184568881988525, + 0.02606494352221489 + ], + "credit_map": [ + 0.030349869281053543, + 0.02222336083650589, + 0.02395310252904892, + 0.015300828963518143, + 0.02198438160121441, + 0.028244879096746445, + 0.024336721748113632, + 0.030141323804855347, + 0.016270577907562256, + 0.008967331610620022, + 0.018293624743819237, + 0.012608972378075123, + 0.0109693743288517, + 0.018441928550601006, + 0.01420333981513977, + 0.015650786459445953, + 0.06696560978889465, + 0.010175292380154133, + 0.013629568740725517, + 0.00810433179140091, + 0.006430752109736204, + 0.07223836332559586, + 0.009170406498014927, + 0.011517329141497612, + 0.0023006501141935587, + 0.0019666145090013742, + 0.002012226264923811, + 0.0017487922450527549, + 0.001083552953787148, + 0.0025742684956640005, + 0.002414500340819359, + 0.0014883017865940928, + 0.03886173292994499, + 0.02411631867289543, + 0.08657568693161011, + 0.017574124038219452, + 0.0185228381305933, + 0.03823148459196091, + 0.034889351576566696, + 0.03062930516898632, + 0.002849637996405363, + 0.002546117641031742, + 0.0037460895255208015, + 0.0025189772713929415, + 0.0010905653471127152, + 0.0024781920947134495, + 0.0015579158207401633, + 0.0016505682142451406, + 0.00542034674435854, + 0.005564691033214331, + 0.0033222679048776627, + 0.005068053025752306, + 0.0022215088829398155, + 0.006104434374719858, + 0.0027953649405390024, + 0.004668044857680798, + 0.006743163336068392, + 0.007577213924378157, + 0.006869789678603411, + 0.009490527212619781, + 0.004134753253310919, + 0.006954630371183157, + 0.005750191863626242, + 0.00446568476036191, + 0.009765304625034332, + 0.006897561252117157, + 0.004623658489435911, + 0.008638161234557629, + 0.003957479260861874, + 0.006260920315980911, + 0.008216668851673603, + 0.0048248195089399815, + 0.00363186770118773, + 0.0038627446629107, + 0.0029719327576458454, + 0.00414120452478528, + 0.002195744775235653, + 0.004047451540827751, + 0.0029772149864584208, + 0.00223678071051836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -1.315321922302246 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743150472640991 + ], + "opposite_logprob_tokens": [ + -0.4743150770664215 + ], + "logprob_reward": -0.9743150472640991, + "yes_no_margin_reward": -0.4999999701976776, + "token_credit_reward": -0.9743150472640991, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.23164311051368713, + 0.057402439415454865, + 0.21911576390266418, + 0.020784122869372368, + 0.21757344901561737, + 0.033684369176626205, + 0.03629550337791443, + 0.034609436988830566, + 0.09280528873205185, + 0.05608652904629707 + ], + "credit_map": [ + 0.03793726861476898, + 0.02703901007771492, + 0.030030928552150726, + 0.020281871780753136, + 0.02684769406914711, + 0.030819015577435493, + 0.027607807889580727, + 0.031079504638910294, + 0.008043751120567322, + 0.006396235432475805, + 0.008604914881289005, + 0.006195636931806803, + 0.006858122535049915, + 0.00799420103430748, + 0.00594608299434185, + 0.007363494485616684, + 0.07112614065408707, + 0.013950652442872524, + 0.014544687233865261, + 0.009939922951161861, + 0.007842188701033592, + 0.07515543699264526, + 0.01232844777405262, + 0.01422828808426857, + 0.0030112124513834715, + 0.0023490360472351313, + 0.0032681527081876993, + 0.0026531852781772614, + 0.0020534247159957886, + 0.002475675893947482, + 0.0026173419319093227, + 0.002356092445552349, + 0.028664471581578255, + 0.019844932481646538, + 0.057858213782310486, + 0.014447120018303394, + 0.014383606612682343, + 0.03095621056854725, + 0.0305156372487545, + 0.02090325951576233, + 0.005141910165548325, + 0.0034818833228200674, + 0.007409204728901386, + 0.003347354708239436, + 0.0024074583780020475, + 0.005819625686854124, + 0.002815868705511093, + 0.00326106371358037, + 0.005065054167062044, + 0.005733970087021589, + 0.004034881480038166, + 0.006754765752702951, + 0.0021133869886398315, + 0.003868014318868518, + 0.0053922864608466625, + 0.0033331431914120913, + 0.004651903174817562, + 0.004558578599244356, + 0.00442184554412961, + 0.00400156993418932, + 0.003273606998845935, + 0.0050938804633915424, + 0.005338229704648256, + 0.0032698241993784904, + 0.01670193299651146, + 0.010244407691061497, + 0.00978401955217123, + 0.013414022512733936, + 0.008431341499090195, + 0.012142549268901348, + 0.012031173333525658, + 0.010055840946733952, + 0.009620591066777706, + 0.005649781785905361, + 0.005309602711349726, + 0.008694615215063095, + 0.005040114279836416, + 0.011158274486660957, + 0.0066610039211809635, + 0.003952545579522848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.9743150472640991 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005961020477116108 + ], + "opposite_logprob_tokens": [ + -5.130960941314697 + ], + "logprob_reward": -0.005961020477116108, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005961020477116108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18010951578617096, + 0.06315337866544724, + 0.255888968706131, + 0.029514271765947342, + 0.20290733873844147, + 0.07646831125020981, + 0.05762484297156334, + 0.04361039400100708, + 0.049113478511571884, + 0.04160946607589722 + ], + "credit_map": [ + 0.02628370001912117, + 0.017344798892736435, + 0.020810309797525406, + 0.015126053243875504, + 0.017956631258130074, + 0.02523503638803959, + 0.025956977158784866, + 0.031396012753248215, + 0.009331351146101952, + 0.0062396340072155, + 0.00865677185356617, + 0.005601933225989342, + 0.006504828575998545, + 0.009593591094017029, + 0.008641763590276241, + 0.00858350284397602, + 0.08309057354927063, + 0.01565445587038994, + 0.01829422079026699, + 0.012365787290036678, + 0.010203266516327858, + 0.08775821328163147, + 0.012669975869357586, + 0.015852484852075577, + 0.00496404804289341, + 0.0034944191575050354, + 0.005077533423900604, + 0.002985343337059021, + 0.0015483532333746552, + 0.005214035511016846, + 0.00321676442399621, + 0.003013774985447526, + 0.027173804119229317, + 0.016570772975683212, + 0.061486147344112396, + 0.01237968448549509, + 0.01257864385843277, + 0.028021346777677536, + 0.024401893839240074, + 0.020295053720474243, + 0.011995258741080761, + 0.014266648329794407, + 0.008753702044487, + 0.008603616617619991, + 0.008302262984216213, + 0.010088243521749973, + 0.007105027325451374, + 0.0073535507544875145, + 0.00907016359269619, + 0.012216426432132721, + 0.004903430584818125, + 0.008430772460997105, + 0.005195796024054289, + 0.007326931692659855, + 0.005684901960194111, + 0.004796420689672232, + 0.00560364592820406, + 0.00594036839902401, + 0.006172363180667162, + 0.005509350448846817, + 0.004108849912881851, + 0.006212056614458561, + 0.006460243370383978, + 0.003603515913709998, + 0.009111269377171993, + 0.006276596337556839, + 0.00455179950222373, + 0.007695849984884262, + 0.0034001250751316547, + 0.005926575977355242, + 0.007537643890827894, + 0.00461361650377512, + 0.005873631685972214, + 0.005629106890410185, + 0.0041728937067091465, + 0.007517338264733553, + 0.0035143897403031588, + 0.005994971375912428, + 0.005414353217929602, + 0.0034927823580801487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.005961020477116108 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048673514276742935 + ], + "opposite_logprob_tokens": [ + -3.048673629760742 + ], + "logprob_reward": -0.048673514276742935, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048673514276742935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2193610668182373, + 0.19467061758041382, + 0.10720697045326233, + 0.2166227251291275, + 0.1291501522064209, + 0.03821847587823868, + 0.024232100695371628, + 0.01807919517159462, + 0.033270128071308136, + 0.01918851211667061 + ], + "credit_map": [ + 0.03871545195579529, + 0.0200053621083498, + 0.026765508577227592, + 0.016603851690888405, + 0.01974673569202423, + 0.03130689635872841, + 0.034042663872241974, + 0.03217461705207825, + 0.025473784655332565, + 0.020314401015639305, + 0.031028419733047485, + 0.018759246915578842, + 0.021977083757519722, + 0.025748247280716896, + 0.021799730136990547, + 0.02956971526145935, + 0.028316911309957504, + 0.005594693589955568, + 0.013839111663401127, + 0.0059097567573189735, + 0.003575757844373584, + 0.029863160103559494, + 0.008204374462366104, + 0.011903214268386364, + 0.03882695361971855, + 0.021220551803708076, + 0.04754522815346718, + 0.01728471741080284, + 0.010128333233296871, + 0.04128158092498779, + 0.01733865961432457, + 0.02299671433866024, + 0.01718216761946678, + 0.011424634605646133, + 0.0363621786236763, + 0.009104734286665916, + 0.009307020343840122, + 0.022258957847952843, + 0.009890170767903328, + 0.01362029742449522, + 0.005706849042326212, + 0.0041282582096755505, + 0.009229857474565506, + 0.00323517550714314, + 0.00273353629745543, + 0.0059105767868459225, + 0.003154856152832508, + 0.004119365010410547, + 0.0036319757346063852, + 0.004822706803679466, + 0.0024757222272455692, + 0.0033593536354601383, + 0.002014934318140149, + 0.0032843167427927256, + 0.0024378548841923475, + 0.002205237979069352, + 0.0025044772773981094, + 0.0019756092224270105, + 0.002261490561068058, + 0.002270650817081332, + 0.001561027835123241, + 0.002578246407210827, + 0.0032225719187408686, + 0.0017051224131137133, + 0.005588148254901171, + 0.00405370257794857, + 0.0032309796661138535, + 0.004759346600621939, + 0.002644933992996812, + 0.004016275051981211, + 0.005857130046933889, + 0.003119614440947771, + 0.002877891529351473, + 0.0029681073501706123, + 0.0017880438826978207, + 0.001860102522186935, + 0.0017354765441268682, + 0.0035383671056479216, + 0.0029767274390906096, + 0.0014437950449064374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.048673514276742935 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298153549432755 + ], + "opposite_logprob_tokens": [ + -2.2379815578460693 + ], + "logprob_reward": -0.11298153549432755, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298153549432755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6465582847595215, + 0.034986626356840134, + 0.045215021818876266, + 0.09191994369029999, + 0.022119352594017982, + 0.044802989810705185, + 0.031213313341140747, + 0.0284255500882864, + 0.03466728702187538, + 0.02009163796901703 + ], + "credit_map": [ + 0.11655836552381516, + 0.06009801849722862, + 0.0829421728849411, + 0.04319792240858078, + 0.06277841329574585, + 0.09782083332538605, + 0.09007589519023895, + 0.0930866152048111, + 0.004398363176733255, + 0.00440451642498374, + 0.005277697928249836, + 0.003912315238267183, + 0.004474834073334932, + 0.004404476378113031, + 0.004190228879451752, + 0.003924194723367691, + 0.013177210465073586, + 0.003503898624330759, + 0.004107209388166666, + 0.0024489073548465967, + 0.002111619571223855, + 0.014188718050718307, + 0.002804145682603121, + 0.002873312681913376, + 0.016042770817875862, + 0.008479572832584381, + 0.019338427111506462, + 0.007651404943317175, + 0.005037009250372648, + 0.017503615468740463, + 0.008104623295366764, + 0.009762522764503956, + 0.003277197713032365, + 0.0019481467315927148, + 0.0036106661427766085, + 0.0022167591378092766, + 0.0013910009292885661, + 0.003676197724416852, + 0.003004346275702119, + 0.0029950367752462626, + 0.006748833693563938, + 0.011513875797390938, + 0.004378076642751694, + 0.004589996300637722, + 0.004791843239217997, + 0.0055323028936982155, + 0.00357910362072289, + 0.0036689576227217913, + 0.005179090891033411, + 0.005979153793305159, + 0.002735762158408761, + 0.0027899888809770346, + 0.0026657383423298597, + 0.004917724058032036, + 0.0038241685833781958, + 0.003121686168015003, + 0.0038804884534329176, + 0.0029340723995119333, + 0.004499239381402731, + 0.0038171112537384033, + 0.0022084801457822323, + 0.004425263963639736, + 0.003342478070408106, + 0.003318416653200984, + 0.004634704440832138, + 0.00408948166295886, + 0.0038494537584483624, + 0.00530971447005868, + 0.0037078617606312037, + 0.00398564338684082, + 0.00639046635478735, + 0.0026999646797776222, + 0.0030696841422468424, + 0.0031763832084834576, + 0.0017088735476136208, + 0.0035778111778199673, + 0.0019371208036318421, + 0.0029605913441628218, + 0.0020826447289437056, + 0.0015785280847921968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.11298153549432755 + } + ], + "vlm_decoded_outputs": [ + "No", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic drone shot slowly zooms forward from high above, tilting slightly upward to reveal a vast vista of a deep canyon, stained golden red by the setting afternoon sun. In the canyon's center, a mountainous skeleton of an ancient creature sits hunched in a posture of eternal contemplation. Its massive, rock-like bones are a weathered off-white, covered in deep cracks and mottled with dark moss, its massive spine curved in a mournful arc. The steep, crimson rock walls cast long, deep shadows, partially shrouding the skeleton in a mysterious darkness, while sunlight sharply outlines the contours of its skull and shoulder blades. A strong wind blew steadily and continuously, sweeping up the dry yellow sand on the ground and forming clearly visible sand flows. These sand flows seemed to have life, whistling and flowing unhindered through the abyss-like empty eye sockets of the skull, and steadily pouring out from the wide gaps between the ribs of its magnificent cathedral dome, forming a misty and continuously flowing sand fog on the leeward side of the skeleton. The whole picture showed movement in an extreme silence, full of solemn, desolate and extremely beautiful atmosphere, conveying an epic sense of loneliness spanning the ages.", + "seed": 15131, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_7/index_3508_seed_15131_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.7104045704007149, + "vlm_logprob": -0.7104045704007149, + "vlm_yes_no_margin": 1.35, + "vlm_token_credit": -0.7104045704007149, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.1281886100769043 + ], + "opposite_logprob_tokens": [ + -0.1281886100769043 + ], + "logprob_reward": -2.1281886100769043, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.1281886100769043, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.29776841402053833, + 0.1557687222957611, + 0.04522765055298805, + 0.0494760163128376, + 0.07788380235433578, + 0.03693544492125511, + 0.05042800307273865, + 0.0346904955804348, + 0.1375134289264679, + 0.11430801451206207 + ], + "credit_map": [ + 0.03554951027035713, + 0.018531344830989838, + 0.06174445524811745, + 0.018412386998534203, + 0.02227281592786312, + 0.038884520530700684, + 0.021421195939183235, + 0.08095218241214752, + 0.01699615828692913, + 0.014246135950088501, + 0.016198357567191124, + 0.01893775723874569, + 0.01701003685593605, + 0.018328063189983368, + 0.03029487282037735, + 0.023757359012961388, + 0.006339807529002428, + 0.0034461086615920067, + 0.005381898954510689, + 0.005100257694721222, + 0.0037682082038372755, + 0.0063226548954844475, + 0.005847473628818989, + 0.009021240286529064, + 0.005225318484008312, + 0.0030821021646261215, + 0.006764660589396954, + 0.005506209097802639, + 0.005277036689221859, + 0.006822824943810701, + 0.005557798780500889, + 0.011240066960453987, + 0.009151868522167206, + 0.007233176380395889, + 0.007351149804890156, + 0.00803674478083849, + 0.007703166920691729, + 0.015184810385107994, + 0.0076184808276593685, + 0.015604402869939804, + 0.004704273771494627, + 0.0030578109435737133, + 0.0025872252881526947, + 0.003745152847841382, + 0.004197534639388323, + 0.0071655819192528725, + 0.005064630415290594, + 0.006413232535123825, + 0.005877622403204441, + 0.006321827881038189, + 0.003946307115256786, + 0.006543516181409359, + 0.008378528989851475, + 0.0065445685759186745, + 0.0058144088834524155, + 0.007001224439591169, + 0.0050585912540555, + 0.002191694686189294, + 0.0016519789351150393, + 0.0023840635549277067, + 0.004663519561290741, + 0.007166425231844187, + 0.0034913686104118824, + 0.008082856424152851, + 0.01165924035012722, + 0.005239048507064581, + 0.004731415770947933, + 0.006575135514140129, + 0.010071891359984875, + 0.04604433476924896, + 0.007007448002696037, + 0.04618491232395172, + 0.014294895343482494, + 0.009843659587204456, + 0.01198373269289732, + 0.015211899764835835, + 0.013195662759244442, + 0.019768785685300827, + 0.01505060214549303, + 0.014958781190216541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -2.1281886100769043 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1371990442276 + ], + "opposite_logprob_tokens": [ + -0.3871990442276001 + ], + "logprob_reward": -1.1371990442276, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1371990442276, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10351335257291794, + 0.166714608669281, + 0.04835542291402817, + 0.11168095469474792, + 0.33885133266448975, + 0.03662894666194916, + 0.042343027889728546, + 0.030398445203900337, + 0.03328295052051544, + 0.0882309228181839 + ], + "credit_map": [ + 0.013115937821567059, + 0.009715573862195015, + 0.017985235899686813, + 0.011399609036743641, + 0.008506144396960735, + 0.017795870080590248, + 0.009887436404824257, + 0.015107542276382446, + 0.02014954574406147, + 0.01175166480243206, + 0.023951653391122818, + 0.021708710119128227, + 0.01906793937087059, + 0.021249720826745033, + 0.023937540128827095, + 0.024897843599319458, + 0.0054562995210289955, + 0.0045224991627037525, + 0.005726359318941832, + 0.006374463904649019, + 0.005718258209526539, + 0.005901869386434555, + 0.007775543723255396, + 0.006880130618810654, + 0.01130000688135624, + 0.010185678489506245, + 0.010923977941274643, + 0.012751341797411442, + 0.014914735220372677, + 0.013325444422662258, + 0.019029855728149414, + 0.019249914214015007, + 0.01925486885011196, + 0.013222388923168182, + 0.01916574314236641, + 0.017060864716768265, + 0.021820848807692528, + 0.11115993559360504, + 0.01853482611477375, + 0.11863186955451965, + 0.0037955476436764, + 0.003074228297919035, + 0.002615739358589053, + 0.0035704236943274736, + 0.005803128704428673, + 0.006394664756953716, + 0.005133525934070349, + 0.006241688039153814, + 0.004889107774943113, + 0.003708865726366639, + 0.0031630657613277435, + 0.004146009683609009, + 0.009025818668305874, + 0.006877731531858444, + 0.005071782041341066, + 0.005460647866129875, + 0.004114564973860979, + 0.002444310812279582, + 0.001240682089701295, + 0.0031365640461444855, + 0.004485199227929115, + 0.006299718748778105, + 0.004430142696946859, + 0.004247261676937342, + 0.004325777292251587, + 0.0021004860755056143, + 0.001801845384761691, + 0.002804174553602934, + 0.00407138979062438, + 0.006302657537162304, + 0.003235593670979142, + 0.008641027845442295, + 0.00945888739079237, + 0.007787441834807396, + 0.008552000857889652, + 0.010032083839178085, + 0.013227471150457859, + 0.017602061852812767, + 0.008563396520912647, + 0.013007582165300846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -1.1371990442276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011107398197054863 + ], + "opposite_logprob_tokens": [ + -4.511107444763184 + ], + "logprob_reward": -0.011107398197054863, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011107398197054863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24454958736896515, + 0.10406285524368286, + 0.03218022733926773, + 0.10716404765844345, + 0.2133992314338684, + 0.07258044928312302, + 0.017517294734716415, + 0.03652309998869896, + 0.11082164198160172, + 0.0612015500664711 + ], + "credit_map": [ + 0.03029262088239193, + 0.014446201734244823, + 0.050131164491176605, + 0.01800292357802391, + 0.018498379737138748, + 0.03613079711794853, + 0.02019982598721981, + 0.056847672909498215, + 0.012030322104692459, + 0.013072142377495766, + 0.010914322920143604, + 0.0115170544013381, + 0.011688061989843845, + 0.012787058018147945, + 0.017857054248452187, + 0.014196841046214104, + 0.005631340201944113, + 0.0026016291230916977, + 0.003363355528563261, + 0.003541632555425167, + 0.0032020241487771273, + 0.005247029010206461, + 0.004160950891673565, + 0.004432267975062132, + 0.009833655320107937, + 0.009249993599951267, + 0.009250197559595108, + 0.01287144236266613, + 0.010224050842225552, + 0.01493161916732788, + 0.016055909916758537, + 0.024747183546423912, + 0.01503676362335682, + 0.009789970703423023, + 0.013336585834622383, + 0.011746974661946297, + 0.014662017114460468, + 0.06565950065851212, + 0.011448707431554794, + 0.07171870023012161, + 0.0076198577880859375, + 0.007798833306878805, + 0.006291403900831938, + 0.009119771420955658, + 0.008792336098849773, + 0.011120493523776531, + 0.009279465302824974, + 0.012558291666209698, + 0.0020831897854804993, + 0.0021301794331520796, + 0.0013352029491215944, + 0.0021770698949694633, + 0.0024979349691420794, + 0.002821305999532342, + 0.001987037481740117, + 0.002485376549884677, + 0.004039012361317873, + 0.0023281124886125326, + 0.0017796012107282877, + 0.0029291228856891394, + 0.004758333787322044, + 0.008367285132408142, + 0.0031531574204564095, + 0.009168475866317749, + 0.007731303106993437, + 0.004091605078428984, + 0.0038443799130618572, + 0.00480762030929327, + 0.008505577221512794, + 0.03948220610618591, + 0.004590350203216076, + 0.037768591195344925, + 0.008611075580120087, + 0.005545280408114195, + 0.006994136143475771, + 0.010082446970045567, + 0.006214759312570095, + 0.007698568515479565, + 0.008889076299965382, + 0.007166204042732716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.011107398197054863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023314977064728737 + ], + "opposite_logprob_tokens": [ + -3.773314952850342 + ], + "logprob_reward": -0.023314977064728737, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023314977064728737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21022558212280273, + 0.1578398197889328, + 0.10214970260858536, + 0.0842900425195694, + 0.06898908317089081, + 0.05346057564020157, + 0.02964494377374649, + 0.19328375160694122, + 0.033148735761642456, + 0.06696777045726776 + ], + "credit_map": [ + 0.030757127329707146, + 0.012276770547032356, + 0.03250539302825928, + 0.0160986240953207, + 0.017752323299646378, + 0.049188338220119476, + 0.019314583390951157, + 0.032332438975572586, + 0.018723981454968452, + 0.013860328122973442, + 0.015413015149533749, + 0.0179631095379591, + 0.018585234880447388, + 0.01814483292400837, + 0.0302454624325037, + 0.024903858080506325, + 0.0171640794724226, + 0.006807022728025913, + 0.012124862521886826, + 0.011136023327708244, + 0.007950196973979473, + 0.017753783613443375, + 0.011165759526193142, + 0.01804797723889351, + 0.008941697888076305, + 0.007806280627846718, + 0.009588023647665977, + 0.008180215023458004, + 0.013666026294231415, + 0.011653682217001915, + 0.010991123504936695, + 0.013462993316352367, + 0.008129812777042389, + 0.004329209215939045, + 0.0068979025818407536, + 0.0048237936571240425, + 0.005401564296334982, + 0.015994276851415634, + 0.004837868735194206, + 0.018574653193354607, + 0.007068137638270855, + 0.005530538037419319, + 0.004167231265455484, + 0.006288040429353714, + 0.005692389328032732, + 0.010385631583631039, + 0.007491531316190958, + 0.00683707557618618, + 0.0033888451289385557, + 0.002853823360055685, + 0.0018518053693696856, + 0.0030214462894946337, + 0.004413532093167305, + 0.007158796768635511, + 0.002750915940850973, + 0.004205780103802681, + 0.026529785245656967, + 0.012824835255742073, + 0.008489137515425682, + 0.014205112121999264, + 0.025337858125567436, + 0.052890367805957794, + 0.01565936952829361, + 0.03734729066491127, + 0.004232182167470455, + 0.0023247122298926115, + 0.0024665878154337406, + 0.003787366207689047, + 0.004694993142038584, + 0.007321120239794254, + 0.004285380244255066, + 0.004036394413560629, + 0.010018350556492805, + 0.005305557977408171, + 0.005298937205225229, + 0.006942277308553457, + 0.009203875437378883, + 0.013160784728825092, + 0.007300260942429304, + 0.00973772443830967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.023314977064728737 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a drone view that slowly zooms forward from high above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt slightly upward to reveal a vast canyon vista?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a mountainous skeleton of an ancient creature visible in the center of the canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the rock walls of the canyon crimson and casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there clearly visible sand flows moving through the eye sockets and rib gaps of the skeleton?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2522128224372864 + ], + "opposite_logprob_tokens": [ + -1.5022127628326416 + ], + "logprob_reward": -0.2522128224372864, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2522128224372864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11003381758928299, + 0.09330292046070099, + 0.04612411931157112, + 0.12891066074371338, + 0.0861918032169342, + 0.04857555404305458, + 0.03991536796092987, + 0.04875272884964943, + 0.2729141116142273, + 0.12527893483638763 + ], + "credit_map": [ + 0.013351176865398884, + 0.010299419052898884, + 0.009607858955860138, + 0.010650329291820526, + 0.013299241662025452, + 0.01838560961186886, + 0.012712412513792515, + 0.02172776125371456, + 0.011896789073944092, + 0.0068880594335496426, + 0.015390885062515736, + 0.010397316887974739, + 0.00795054156333208, + 0.012118310667574406, + 0.016159845516085625, + 0.012501159682869911, + 0.005858732387423515, + 0.0033230234403163195, + 0.006113224662840366, + 0.005437961779534817, + 0.005302023608237505, + 0.005831785500049591, + 0.0061899470165371895, + 0.008067414164543152, + 0.01353544369339943, + 0.010779056698083878, + 0.016701865941286087, + 0.013569192960858345, + 0.019188186153769493, + 0.017639385536313057, + 0.01941480115056038, + 0.01808272674679756, + 0.005321217700839043, + 0.004075699485838413, + 0.005168048664927483, + 0.0049120341427624226, + 0.006617386359721422, + 0.02531415969133377, + 0.005678493529558182, + 0.029104746878147125, + 0.004848564974963665, + 0.0030364205595105886, + 0.0035756865981966257, + 0.005676129832863808, + 0.007961004972457886, + 0.008343844674527645, + 0.005179478786885738, + 0.009954418055713177, + 0.0047044409438967705, + 0.003970015328377485, + 0.002327674999833107, + 0.004196255002170801, + 0.007382804993540049, + 0.007421340327709913, + 0.0041658030822873116, + 0.005747027229517698, + 0.005659405142068863, + 0.003932718653231859, + 0.0028653056360781193, + 0.004648719448596239, + 0.008382800966501236, + 0.01179309468716383, + 0.0048453020863235, + 0.0066253733821213245, + 0.02466088905930519, + 0.012422309257090092, + 0.012406229972839355, + 0.017420370131731033, + 0.026088913902640343, + 0.08812537044286728, + 0.016451265662908554, + 0.07533875107765198, + 0.016405608505010605, + 0.009785693138837814, + 0.010886138305068016, + 0.021548589691519737, + 0.014713709242641926, + 0.0172326248139143, + 0.019459566101431847, + 0.015247001312673092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.2522128224372864 + } + ], + "vlm_decoded_outputs": [ + "No", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_8/index_tensor([5621, 5621])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_8/index_tensor([5621, 5621])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3fe2d30d24f83c04758b4f0b3ef958b63a03243b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_8/index_tensor([5621, 5621])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady medium shot, captured from a level perspective, captures a moment steeped in a sense of heritage. Inside a dimly lit, antique-style Chinese medicine shop, an elderly master, his silver hair meticulously combed and his face lined with wisdom, wears a dark blue cotton and linen double-breasted jacket, sleeves slightly rolled up, revealing his veiny yet remarkably strong forearms. He calmly and solemnly hands over a carefully selected bundle of herbs. This bundle is truly extraordinary: a medley of earthy, dark brown roots, a few broad, unfurling dark green leaves, and a handful of small, lavender dried flowers, bound together with a simple hemp string. Across from him, a young apprentice bows slightly, hands respectfully extended, ready to receive. Dressed in a simple off-white, coarse cloth shirt, the apprentice's gaze is focused and reverent, his entire attention focused on his master's hands and the bundle. In the background stood a massive, neatly arranged medicine cabinet, its countless small drawers covered with yellowed paper labels. The air seemed to be filled with the faint aroma of medicine and dust. A ray of afternoon sunlight slanted in through the carved wooden window on the side, forming a sharp beam that perfectly illuminated the two people's joined hands and the bundle of herbs, gilding this tranquil, focused, and ritualistic scene with a warm and sacred golden aura.", + "seed": 21038, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_8/index_5621_seed_21038_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.015341171389445662, + "vlm_logprob": -0.015341171389445662, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.015341171389445662, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03857295215129852 + ], + "opposite_logprob_tokens": [ + -3.9135730266571045 + ], + "logprob_reward": -0.03857295215129852, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03857295215129852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2667914628982544, + 0.07761415839195251, + 0.11210328340530396, + 0.13364443182945251, + 0.0791320949792862, + 0.1512831449508667, + 0.055527061223983765, + 0.04957613721489906, + 0.04537452384829521, + 0.028953738510608673 + ], + "credit_map": [ + 0.025652799755334854, + 0.022447124123573303, + 0.09021865576505661, + 0.03610575944185257, + 0.011916915886104107, + 0.01867079921066761, + 0.03453516215085983, + 0.027244243770837784, + 0.010682731866836548, + 0.007598093245178461, + 0.010755185969173908, + 0.015118423849344254, + 0.006123858503997326, + 0.006537799723446369, + 0.009031706489622593, + 0.01176635641604662, + 0.021737150847911835, + 0.012588169425725937, + 0.016542011871933937, + 0.0172586552798748, + 0.007918552495539188, + 0.009225225076079369, + 0.011493124067783356, + 0.015340386889874935, + 0.01597592607140541, + 0.02616880089044571, + 0.024698598310351372, + 0.020553546026349068, + 0.00781931634992361, + 0.007929776795208454, + 0.008824607357382774, + 0.021673865616321564, + 0.01210465095937252, + 0.010581129230558872, + 0.017560994252562523, + 0.00967524666339159, + 0.007835903204977512, + 0.006295765750110149, + 0.004719509743154049, + 0.010358895175158978, + 0.01952356845140457, + 0.01796604134142399, + 0.04045844078063965, + 0.0167723149061203, + 0.015865743160247803, + 0.009533527307212353, + 0.011306117288768291, + 0.01985739916563034, + 0.00529158441349864, + 0.004520211834460497, + 0.007167911157011986, + 0.011511779390275478, + 0.005191621836274862, + 0.007189481053501368, + 0.006873046513646841, + 0.007781425025314093, + 0.010991545394062996, + 0.006915649864822626, + 0.006938784848898649, + 0.004856306128203869, + 0.005022627301514149, + 0.004374510142952204, + 0.003824731335043907, + 0.006651982665061951, + 0.0048290640115737915, + 0.006918956059962511, + 0.004949870985001326, + 0.00482107512652874, + 0.007182522211223841, + 0.005544483661651611, + 0.004590783733874559, + 0.006537765730172396, + 0.002463898854330182, + 0.004076871555298567, + 0.005838168319314718, + 0.0039028641767799854, + 0.0026793808210641146, + 0.0031349025666713715, + 0.0031343346927314997, + 0.003723317291587591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.03857295215129852 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002516676438972354 + ], + "opposite_logprob_tokens": [ + -6.002516746520996 + ], + "logprob_reward": -0.002516676438972354, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002516676438972354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31571561098098755, + 0.06190892308950424, + 0.16390188038349152, + 0.09080082923173904, + 0.052868522703647614, + 0.08520405739545822, + 0.1081971749663353, + 0.05690842494368553, + 0.032852672040462494, + 0.0316418781876564 + ], + "credit_map": [ + 0.021915854886174202, + 0.021141909062862396, + 0.11845533549785614, + 0.04119883105158806, + 0.01615448109805584, + 0.024087145924568176, + 0.025444570928812027, + 0.047317489981651306, + 0.004716240800917149, + 0.005928122438490391, + 0.00821326207369566, + 0.01055973581969738, + 0.0065620350651443005, + 0.009755231440067291, + 0.008952981792390347, + 0.0072213225066661835, + 0.019329305738210678, + 0.012293988838791847, + 0.023026874288916588, + 0.04417372867465019, + 0.010251732543110847, + 0.013609989546239376, + 0.017698341980576515, + 0.023517947643995285, + 0.006813038606196642, + 0.005977720953524113, + 0.019271356984972954, + 0.008225751109421253, + 0.005662872921675444, + 0.018553273752331734, + 0.014015503227710724, + 0.012281312607228756, + 0.004448675084859133, + 0.004322908818721771, + 0.014204651117324829, + 0.006867919582873583, + 0.004251697566360235, + 0.005722996313124895, + 0.005511041264981031, + 0.007538636680692434, + 0.007730206940323114, + 0.01380909513682127, + 0.011149573139846325, + 0.009825420565903187, + 0.014846031554043293, + 0.008743456564843655, + 0.008116728626191616, + 0.010983551852405071, + 0.010822314769029617, + 0.021352842450141907, + 0.011105946265161037, + 0.01372579950839281, + 0.014449005015194416, + 0.011205939576029778, + 0.008554748259484768, + 0.016980590298771858, + 0.009197019971907139, + 0.009906475432217121, + 0.008383267559111118, + 0.006555669009685516, + 0.005949518643319607, + 0.005066575016826391, + 0.004007054027169943, + 0.00784284807741642, + 0.0029915974009782076, + 0.006721969228237867, + 0.003219427540898323, + 0.003876105882227421, + 0.0027713270392268896, + 0.003674468258395791, + 0.005021192599087954, + 0.004576585255563259, + 0.002471447689458728, + 0.005926924291998148, + 0.004046885296702385, + 0.0038695703260600567, + 0.0030341895762830973, + 0.004498905967921019, + 0.0035321861505508423, + 0.0042617712169885635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.002516676438972354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003623825730755925 + ], + "opposite_logprob_tokens": [ + -5.628623962402344 + ], + "logprob_reward": -0.003623825730755925, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003623825730755925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19605301320552826, + 0.14118675887584686, + 0.09973061084747314, + 0.06981422007083893, + 0.10968511551618576, + 0.08588238060474396, + 0.1120709776878357, + 0.08806097507476807, + 0.05285034328699112, + 0.044665660709142685 + ], + "credit_map": [ + 0.01564885303378105, + 0.011988860554993153, + 0.052055828273296356, + 0.02967468649148941, + 0.013547475449740887, + 0.024146156385540962, + 0.025050034746527672, + 0.023941123858094215, + 0.007893090136349201, + 0.012088836170732975, + 0.026057472452521324, + 0.03315233439207077, + 0.008995546028017998, + 0.01688573509454727, + 0.019205793738365173, + 0.016907954588532448, + 0.017746858298778534, + 0.010247278958559036, + 0.01339372992515564, + 0.01402345672249794, + 0.007173136807978153, + 0.008940493687987328, + 0.01467157807201147, + 0.013534082099795341, + 0.008993759751319885, + 0.010016187094151974, + 0.01077481359243393, + 0.011148682795464993, + 0.004836352076381445, + 0.007215746212750673, + 0.007445862982422113, + 0.009382818825542927, + 0.013884500600397587, + 0.009191486053168774, + 0.035728875547647476, + 0.012618949636816978, + 0.00945561658591032, + 0.008360489271581173, + 0.006484679877758026, + 0.013960512354969978, + 0.011750845238566399, + 0.011687627993524075, + 0.02024812251329422, + 0.009558211080729961, + 0.007268156856298447, + 0.007589448243379593, + 0.0071207680739462376, + 0.010659204795956612, + 0.011699254624545574, + 0.010122797451913357, + 0.016543414443731308, + 0.024729976430535316, + 0.007400304079055786, + 0.01208712812513113, + 0.011908300220966339, + 0.017579801380634308, + 0.010604283772408962, + 0.016481459140777588, + 0.013669637031853199, + 0.00917512085288763, + 0.01099733728915453, + 0.006912633311003447, + 0.007743015885353088, + 0.012477485463023186, + 0.004314472898840904, + 0.00940499734133482, + 0.0053845555521547794, + 0.006791913881897926, + 0.006353176664561033, + 0.0069342791102826595, + 0.006009552627801895, + 0.00765739381313324, + 0.0036896381061524153, + 0.005971385631710291, + 0.0050323037430644035, + 0.0070818630047142506, + 0.00638169189915061, + 0.006498661357909441, + 0.0037743858993053436, + 0.006235731299966574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.003623825730755925 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021961636375635862 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021961636375635862, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021961636375635862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28793781995773315, + 0.03677188232541084, + 0.04574234038591385, + 0.042776186019182205, + 0.06320199370384216, + 0.06252799928188324, + 0.26203614473342896, + 0.08216060698032379, + 0.0728684589266777, + 0.043976642191410065 + ], + "credit_map": [ + 0.021125512197613716, + 0.01888749562203884, + 0.10511856526136398, + 0.03681310638785362, + 0.0138286417350173, + 0.022323496639728546, + 0.03773074224591255, + 0.032110244035720825, + 0.0028956737369298935, + 0.004091594833880663, + 0.004819259513169527, + 0.007602676283568144, + 0.0028744805604219437, + 0.004269629716873169, + 0.0053986841812729836, + 0.004819884430617094, + 0.005404868628829718, + 0.0031446130014955997, + 0.005551757290959358, + 0.008500236086547375, + 0.0033107539638876915, + 0.006142078433185816, + 0.006887116935104132, + 0.006800915114581585, + 0.0038441510405391455, + 0.005641909781843424, + 0.009149772115051746, + 0.004477315116673708, + 0.002776626031845808, + 0.0066187744960188866, + 0.004415000788867474, + 0.0058526345528662205, + 0.007283487822860479, + 0.006721015553921461, + 0.01971697434782982, + 0.0067986296489834785, + 0.004370294976979494, + 0.006416265852749348, + 0.0038778423331677914, + 0.008017478510737419, + 0.008574421517550945, + 0.008411657065153122, + 0.010709359310567379, + 0.008899811655282974, + 0.006002521142363548, + 0.005494521930813789, + 0.006344831082969904, + 0.008090878836810589, + 0.027565432712435722, + 0.04007973521947861, + 0.030329087749123573, + 0.04965599626302719, + 0.025446105748414993, + 0.020660901442170143, + 0.028488509356975555, + 0.039810385555028915, + 0.008936122059822083, + 0.010377483442425728, + 0.013921838253736496, + 0.01759556122124195, + 0.0057452027685940266, + 0.006293909158557653, + 0.007913798093795776, + 0.011376684531569481, + 0.007083286065608263, + 0.012071254663169384, + 0.0067156036384403706, + 0.011533321812748909, + 0.006942203734070063, + 0.008369780145585537, + 0.00944762397557497, + 0.010705380700528622, + 0.003914909902960062, + 0.006702046375721693, + 0.004666920285671949, + 0.010636705905199051, + 0.0036269875708967447, + 0.005102380178868771, + 0.003842218080535531, + 0.005484479013830423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0021961636375635862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796238988637924 + ], + "opposite_logprob_tokens": [ + -3.5297961235046387 + ], + "logprob_reward": -0.029796238988637924, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796238988637924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5110974907875061, + 0.06299711018800735, + 0.05714916065335274, + 0.04428107291460037, + 0.0399833582341671, + 0.06693714112043381, + 0.10662692785263062, + 0.055824197828769684, + 0.030979814007878304, + 0.02412375621497631 + ], + "credit_map": [ + 0.03282411769032478, + 0.03535730764269829, + 0.0901825800538063, + 0.07358288019895554, + 0.05063077062368393, + 0.07653412967920303, + 0.07972735166549683, + 0.072258360683918, + 0.006047114264219999, + 0.006427037063986063, + 0.007024986203759909, + 0.01433408260345459, + 0.003353985259309411, + 0.005207340233027935, + 0.011945599690079689, + 0.008656964637339115, + 0.0041922652162611485, + 0.006348187103867531, + 0.008989616297185421, + 0.009759696200489998, + 0.005747655872255564, + 0.006128779612481594, + 0.006917337886989117, + 0.00906562153249979, + 0.005574606824666262, + 0.0051466855220496655, + 0.008432199247181416, + 0.006193625275045633, + 0.002674652496352792, + 0.003620060160756111, + 0.007254176307469606, + 0.005385067313909531, + 0.005324327852576971, + 0.00471371877938509, + 0.007767968345433474, + 0.005978131666779518, + 0.00375222391448915, + 0.0032264532055705786, + 0.003601998556405306, + 0.005618535913527012, + 0.006552070379257202, + 0.007043932564556599, + 0.014949985779821873, + 0.009172718040645123, + 0.005446215160191059, + 0.008543364703655243, + 0.006979974452406168, + 0.008248881436884403, + 0.014153754338622093, + 0.021382249891757965, + 0.013719236478209496, + 0.011948742903769016, + 0.01327096950262785, + 0.008033341728150845, + 0.00828670896589756, + 0.015831924974918365, + 0.005337273236364126, + 0.009701695293188095, + 0.006974135059863329, + 0.009053359739482403, + 0.007199692539870739, + 0.005289821419864893, + 0.00463070347905159, + 0.007637514732778072, + 0.0027502074372023344, + 0.005485134664922953, + 0.00408485671505332, + 0.002987754065543413, + 0.00362818269059062, + 0.004185468889772892, + 0.002864504000172019, + 0.004993705544620752, + 0.0020111582707613707, + 0.0040882364846765995, + 0.0036259593907743692, + 0.0033502730075269938, + 0.002291201613843441, + 0.002405093749985099, + 0.002733467146754265, + 0.0036183674819767475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.029796238988637924 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, captured from a level perspective, captures a moment steeped in a sense of heritage. Inside a dimly lit, antique-style Chinese medicine shop, an elderly master, his silver hair meticulously combed and his face lined with wisdom, wears a dark blue cotton and linen double-breasted jacket, sleeves slightly rolled up, revealing his veiny yet remarkably strong forearms. He calmly and solemnly hands over a carefully selected bundle of herbs. This bundle is truly extraordinary: a medley of earthy, dark brown roots, a few broad, unfurling dark green leaves, and a handful of small, lavender dried flowers, bound together with a simple hemp string. Across from him, a young apprentice bows slightly, hands respectfully extended, ready to receive. Dressed in a simple off-white, coarse cloth shirt, the apprentice's gaze is focused and reverent, his entire attention focused on his master's hands and the bundle. In the background stood a massive, neatly arranged medicine cabinet, its countless small drawers covered with yellowed paper labels. The air seemed to be filled with the faint aroma of medicine and dust. A ray of afternoon sunlight slanted in through the carved wooden window on the side, forming a sharp beam that perfectly illuminated the two people's joined hands and the bundle of herbs, gilding this tranquil, focused, and ritualistic scene with a warm and sacred golden aura.", + "seed": 21039, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_8/index_5621_seed_21039_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.011883279029279947, + "vlm_logprob": -0.011883279029279947, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.011883279029279947, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02197802998125553 + ], + "opposite_logprob_tokens": [ + -4.14697790145874 + ], + "logprob_reward": -0.02197802998125553, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02197802998125553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2449195384979248, + 0.07790294289588928, + 0.03169003501534462, + 0.2868589460849762, + 0.0725250318646431, + 0.07883074879646301, + 0.04878613352775574, + 0.013120084069669247, + 0.016326019540429115, + 0.12904050946235657 + ], + "credit_map": [ + 0.025201337411999702, + 0.02324194274842739, + 0.02524440549314022, + 0.013063115999102592, + 0.035444047302007675, + 0.021573161706328392, + 0.03487473353743553, + 0.06627678871154785, + 0.0053484211675822735, + 0.009056122042238712, + 0.012789394706487656, + 0.007686762139201164, + 0.01848454214632511, + 0.007124120369553566, + 0.010731921531260014, + 0.006681662052869797, + 0.0028790857177227736, + 0.005651939660310745, + 0.003634369233623147, + 0.004319788422435522, + 0.005035557318478823, + 0.003057928057387471, + 0.005280923563987017, + 0.0018304437398910522, + 0.025438571348786354, + 0.032298192381858826, + 0.023117590695619583, + 0.013515261001884937, + 0.032378122210502625, + 0.01568877138197422, + 0.10500596463680267, + 0.03941645100712776, + 0.0148762296885252, + 0.009100659750401974, + 0.011838246136903763, + 0.007117696106433868, + 0.009983508847653866, + 0.007073635701090097, + 0.007577978540211916, + 0.004957079887390137, + 0.006051083095371723, + 0.004229995887726545, + 0.015263479202985764, + 0.006566715892404318, + 0.005878448020666838, + 0.005632871761918068, + 0.025622615590691566, + 0.009585543535649776, + 0.0034808660857379436, + 0.00463893823325634, + 0.00831366516649723, + 0.0038667097687721252, + 0.007192250341176987, + 0.008687074296176434, + 0.00464262068271637, + 0.00796401035040617, + 0.00116376590449363, + 0.0015203236835077405, + 0.0025211197789758444, + 0.001897522946819663, + 0.0015378237003460526, + 0.0015619121259078383, + 0.0012474855175241828, + 0.0016701299464330077, + 0.0023045081179589033, + 0.0016528096748515964, + 0.0031665179412811995, + 0.002074686111882329, + 0.0018474849639460444, + 0.0010655589867383242, + 0.0015193934086710215, + 0.0026950612664222717, + 0.013376918621361256, + 0.005240297876298428, + 0.038078099489212036, + 0.012086891569197178, + 0.008869430050253868, + 0.008861531503498554, + 0.0045145172625780106, + 0.03801283240318298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.02197802998125553 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028342336881905794 + ], + "opposite_logprob_tokens": [ + -5.877834320068359 + ], + "logprob_reward": -0.0028342336881905794, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028342336881905794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27263131737709045, + 0.1639212667942047, + 0.021628262475132942, + 0.14400412142276764, + 0.04270613566040993, + 0.0442625917494297, + 0.11753927171230316, + 0.012412487529218197, + 0.05508610978722572, + 0.12580843269824982 + ], + "credit_map": [ + 0.01882638782262802, + 0.052818816155195236, + 0.02414705976843834, + 0.028614584356546402, + 0.04966631531715393, + 0.02760501019656658, + 0.04741685464978218, + 0.023536304011940956, + 0.014266831800341606, + 0.03042392060160637, + 0.02033967338502407, + 0.01784808747470379, + 0.035185307264328, + 0.014567517675459385, + 0.023923015221953392, + 0.007366914302110672, + 0.0034106778912246227, + 0.0024042311124503613, + 0.0020854163449257612, + 0.0030384049750864506, + 0.003235580399632454, + 0.0027797326911240816, + 0.0029436710756272078, + 0.001730548799969256, + 0.011016396805644035, + 0.015420990996062756, + 0.011155841872096062, + 0.005810968577861786, + 0.01858077570796013, + 0.009025709703564644, + 0.052103251218795776, + 0.020890183746814728, + 0.0046082716435194016, + 0.006848029792308807, + 0.006726051680743694, + 0.00659887446090579, + 0.0056401751935482025, + 0.004612302407622337, + 0.00532898074015975, + 0.0023434485774487257, + 0.0037561566568911076, + 0.0020211003720760345, + 0.006235828623175621, + 0.003664979711174965, + 0.0028948625549674034, + 0.0031299740076065063, + 0.016064004972577095, + 0.006495685316622257, + 0.009004123508930206, + 0.007578574120998383, + 0.017634162679314613, + 0.009674226865172386, + 0.011784644797444344, + 0.027236105874180794, + 0.010234881192445755, + 0.02439255639910698, + 0.0009218972409144044, + 0.0013998441863805056, + 0.0021139010787010193, + 0.0016088441479951143, + 0.0017402798403054476, + 0.001913484651595354, + 0.0013455518055707216, + 0.0013686843449249864, + 0.0067652142606675625, + 0.005058890208601952, + 0.012001406401395798, + 0.00746532529592514, + 0.00429087970405817, + 0.005672224331647158, + 0.0043720221146941185, + 0.0094601446762681, + 0.013417894020676613, + 0.004522308707237244, + 0.037123799324035645, + 0.01102938037365675, + 0.00823190063238144, + 0.008698639459908009, + 0.005133440252393484, + 0.03765105828642845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0028342336881905794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036219253670424223 + ], + "opposite_logprob_tokens": [ + -5.628622055053711 + ], + "logprob_reward": -0.0036219253670424223, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036219253670424223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08106749504804611, + 0.13269655406475067, + 0.04803057387471199, + 0.12877120077610016, + 0.0361969918012619, + 0.27175772190093994, + 0.15017233788967133, + 0.04528915882110596, + 0.03584262728691101, + 0.07017534971237183 + ], + "credit_map": [ + 0.009331588633358479, + 0.00800179410725832, + 0.009356142953038216, + 0.00470153521746397, + 0.010927333496510983, + 0.007839441299438477, + 0.008482728153467178, + 0.02242692932486534, + 0.013819015584886074, + 0.017033524811267853, + 0.01324267964810133, + 0.01098297443240881, + 0.03562556207180023, + 0.0131630664691329, + 0.02121562696993351, + 0.00761408731341362, + 0.006372073665261269, + 0.010118301957845688, + 0.0037806991022080183, + 0.005352849140763283, + 0.006760010030120611, + 0.0035298806615173817, + 0.006267460063099861, + 0.005849296692758799, + 0.009391413070261478, + 0.013622701168060303, + 0.010408159345388412, + 0.007567875552922487, + 0.013457760214805603, + 0.00851428136229515, + 0.04613805189728737, + 0.019670959562063217, + 0.003260234370827675, + 0.0060730925761163235, + 0.005736345425248146, + 0.004646469838917255, + 0.005283184349536896, + 0.0042039030231535435, + 0.004700486082583666, + 0.002293272642418742, + 0.018424034118652344, + 0.015538466162979603, + 0.051040299236774445, + 0.033925991505384445, + 0.019047411158680916, + 0.017931131646037102, + 0.08799698203802109, + 0.027853380888700485, + 0.0156285148113966, + 0.016203545033931732, + 0.02679443359375, + 0.0247690137475729, + 0.02304852195084095, + 0.013767840340733528, + 0.016316762194037437, + 0.013643705286085606, + 0.004602080676704645, + 0.0066433618776500225, + 0.007679591421037912, + 0.007050570100545883, + 0.004921178333461285, + 0.00487732095643878, + 0.005276455543935299, + 0.004238597117364407, + 0.00828320812433958, + 0.0028854927513748407, + 0.006703286897391081, + 0.004207685589790344, + 0.00452836649492383, + 0.0026406666729599237, + 0.003746839240193367, + 0.0028470803517848253, + 0.008416460826992989, + 0.005378527566790581, + 0.013781968504190445, + 0.011560904793441296, + 0.00617666682228446, + 0.005792063195258379, + 0.00801759585738182, + 0.011051155626773834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.0036219253670424223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635777790099382 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004635777790099382, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635777790099382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13192839920520782, + 0.1744467169046402, + 0.020328665152192116, + 0.02095690742135048, + 0.04372474178671837, + 0.12401071190834045, + 0.33946293592453003, + 0.012521175667643547, + 0.018590711057186127, + 0.11402902007102966 + ], + "credit_map": [ + 0.012828162871301174, + 0.011807285249233246, + 0.01583879627287388, + 0.007282057777047157, + 0.024341175332665443, + 0.012145220302045345, + 0.017017938196659088, + 0.030667761340737343, + 0.017370643094182014, + 0.025287799537181854, + 0.016918689012527466, + 0.01559055969119072, + 0.04448353871703148, + 0.01714717224240303, + 0.025521470233798027, + 0.012126852758228779, + 0.0026715919375419617, + 0.003347699763253331, + 0.001793055678717792, + 0.0024242878425866365, + 0.003170178271830082, + 0.0024450013879686594, + 0.002558779902756214, + 0.0019180721137672663, + 0.0043455082923173904, + 0.0027758560609072447, + 0.002057470381259918, + 0.0034858633298426867, + 0.002524212934076786, + 0.001660551642999053, + 0.0027112863026559353, + 0.0013961601071059704, + 0.004420396406203508, + 0.008676656521856785, + 0.005816490855067968, + 0.005553778726607561, + 0.0071063474752008915, + 0.005047219805419445, + 0.004595918580889702, + 0.0025079387705773115, + 0.009690695442259312, + 0.005175378173589706, + 0.01814802549779415, + 0.008383362554013729, + 0.007888495922088623, + 0.009139426052570343, + 0.04590609297156334, + 0.019679240882396698, + 0.026343150064349174, + 0.02267712540924549, + 0.05126800015568733, + 0.026871241629123688, + 0.03406042978167534, + 0.0781383216381073, + 0.029661227017641068, + 0.07044348865747452, + 0.0011072041234001517, + 0.0018845839658752084, + 0.0020074457861483097, + 0.0013511269353330135, + 0.0015970009844750166, + 0.0017637148266658187, + 0.001372440019622445, + 0.001437659258954227, + 0.0024645025841891766, + 0.001784625230357051, + 0.00390159641392529, + 0.001953979954123497, + 0.0020676073618233204, + 0.0012240023352205753, + 0.0015241596847772598, + 0.003670238424092531, + 0.012648158706724644, + 0.004151643253862858, + 0.033374570310115814, + 0.010955221019685268, + 0.007449025753885508, + 0.0075430115684866905, + 0.004050782416015863, + 0.03385661542415619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.004635777790099382 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634642831981182 + ], + "opposite_logprob_tokens": [ + -3.651346445083618 + ], + "logprob_reward": -0.02634642831981182, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634642831981182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07047281414270401, + 0.11949068307876587, + 0.030885523185133934, + 0.07357365638017654, + 0.46710440516471863, + 0.09841851145029068, + 0.06245807558298111, + 0.024248946458101273, + 0.018923604860901833, + 0.0344238318502903 + ], + "credit_map": [ + 0.008831735700368881, + 0.0075005777180194855, + 0.009599138982594013, + 0.006206072401255369, + 0.008813594467937946, + 0.006414829753339291, + 0.00863197073340416, + 0.014474889263510704, + 0.011320794001221657, + 0.015137183479964733, + 0.01562105305492878, + 0.010378403589129448, + 0.026130348443984985, + 0.013912924565374851, + 0.016989165917038918, + 0.010000806301832199, + 0.003793778596445918, + 0.005697408691048622, + 0.003063637064769864, + 0.0030895534437149763, + 0.00512693403288722, + 0.0021990351378917694, + 0.004270327743142843, + 0.0036448489408940077, + 0.006958156358450651, + 0.008658098988234997, + 0.006150327622890472, + 0.006671976763755083, + 0.011050854809582233, + 0.006039989180862904, + 0.020355576649308205, + 0.007688675541430712, + 0.05330544337630272, + 0.09323474764823914, + 0.06503033638000488, + 0.055588994175195694, + 0.07902595400810242, + 0.04298801347613335, + 0.051940545439720154, + 0.025990379974246025, + 0.0072729079984128475, + 0.005931127816438675, + 0.020671067759394646, + 0.010891183279454708, + 0.007952450774610043, + 0.006497830152511597, + 0.02970227226614952, + 0.009499672800302505, + 0.004476055037230253, + 0.005925242323428392, + 0.009731623344123363, + 0.006316549144685268, + 0.00792427733540535, + 0.01196777168661356, + 0.004659654572606087, + 0.011456904001533985, + 0.002505943179130554, + 0.0033247927203774452, + 0.004551962483674288, + 0.002668345347046852, + 0.002948000328615308, + 0.0025732950307428837, + 0.0021030656062066555, + 0.0035735429264605045, + 0.002847878960892558, + 0.0015466171316802502, + 0.0038808248937129974, + 0.003295206930488348, + 0.0018027520272880793, + 0.0014453891199082136, + 0.0019099481869488955, + 0.002194988075643778, + 0.003527113702148199, + 0.003956255968660116, + 0.007372972089797258, + 0.004382013343274593, + 0.0031257611699402332, + 0.003116545733064413, + 0.0026433542370796204, + 0.006299817468971014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.02634642831981182 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, captured from a level perspective, captures a moment steeped in a sense of heritage. Inside a dimly lit, antique-style Chinese medicine shop, an elderly master, his silver hair meticulously combed and his face lined with wisdom, wears a dark blue cotton and linen double-breasted jacket, sleeves slightly rolled up, revealing his veiny yet remarkably strong forearms. He calmly and solemnly hands over a carefully selected bundle of herbs. This bundle is truly extraordinary: a medley of earthy, dark brown roots, a few broad, unfurling dark green leaves, and a handful of small, lavender dried flowers, bound together with a simple hemp string. Across from him, a young apprentice bows slightly, hands respectfully extended, ready to receive. Dressed in a simple off-white, coarse cloth shirt, the apprentice's gaze is focused and reverent, his entire attention focused on his master's hands and the bundle. In the background stood a massive, neatly arranged medicine cabinet, its countless small drawers covered with yellowed paper labels. The air seemed to be filled with the faint aroma of medicine and dust. A ray of afternoon sunlight slanted in through the carved wooden window on the side, forming a sharp beam that perfectly illuminated the two people's joined hands and the bundle of herbs, gilding this tranquil, focused, and ritualistic scene with a warm and sacred golden aura.", + "seed": 21040, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_8/index_5621_seed_21040_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027813691878691316, + "vlm_logprob": -0.027813691878691316, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.027813691878691316, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02708832360804081 + ], + "opposite_logprob_tokens": [ + -4.527088165283203 + ], + "logprob_reward": -0.02708832360804081, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.02708832360804081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.262905091047287, + 0.06257778406143188, + 0.04501703381538391, + 0.040439579635858536, + 0.07760526239871979, + 0.04245470464229584, + 0.04480413347482681, + 0.03837375342845917, + 0.25967973470687866, + 0.1261429488658905 + ], + "credit_map": [ + 0.05446723848581314, + 0.023032840341329575, + 0.034083686769008636, + 0.030861206352710724, + 0.018696388229727745, + 0.028508782386779785, + 0.038911573588848114, + 0.03434339910745621, + 0.010795876383781433, + 0.005473962053656578, + 0.008774785324931145, + 0.00796069111675024, + 0.006609155330806971, + 0.008218512870371342, + 0.007712367456406355, + 0.007032434921711683, + 0.006617818493396044, + 0.005295334849506617, + 0.004754789173603058, + 0.008178578689694405, + 0.004179816693067551, + 0.004588481970131397, + 0.005672444589436054, + 0.0057297684252262115, + 0.005571493413299322, + 0.005040926393121481, + 0.005112417042255402, + 0.005334344692528248, + 0.00413381215184927, + 0.005082887131720781, + 0.004165089223533869, + 0.0059986114501953125, + 0.0087134949862957, + 0.009144606068730354, + 0.008181469514966011, + 0.016023950651288033, + 0.007360554300248623, + 0.014218715950846672, + 0.005243662279099226, + 0.008718805387616158, + 0.007495197933167219, + 0.005989931523799896, + 0.004226143937557936, + 0.005137968342751265, + 0.005591068882495165, + 0.004998670890927315, + 0.0038579548709094524, + 0.005157770123332739, + 0.00686063664034009, + 0.007484854198992252, + 0.00770871527493, + 0.004149433225393295, + 0.004216773901134729, + 0.00503935432061553, + 0.004023912828415632, + 0.005320454016327858, + 0.005455316044390202, + 0.005316001828759909, + 0.004561068490147591, + 0.005130299367010593, + 0.004078004974871874, + 0.005271400325000286, + 0.0035592622589319944, + 0.005002401769161224, + 0.06330839544534683, + 0.02021106518805027, + 0.028413239866495132, + 0.012775732204318047, + 0.01837175153195858, + 0.07377515733242035, + 0.011706147342920303, + 0.031118227168917656, + 0.013620841316878796, + 0.013927035965025425, + 0.03067178651690483, + 0.011974887922406197, + 0.012045984156429768, + 0.019681062549352646, + 0.010942485183477402, + 0.01327886339277029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.02708832360804081 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032165716402232647 + ], + "opposite_logprob_tokens": [ + -5.753216743469238 + ], + "logprob_reward": -0.0032165716402232647, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032165716402232647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2737842798233032, + 0.22124522924423218, + 0.15723001956939697, + 0.03333082050085068, + 0.06377267837524414, + 0.04289049655199051, + 0.04381951317191124, + 0.05232425034046173, + 0.053743816912174225, + 0.05785892531275749 + ], + "credit_map": [ + 0.04137920215725899, + 0.0194607712328434, + 0.04289930313825607, + 0.024597810581326485, + 0.012087090872228146, + 0.0518232025206089, + 0.04237240552902222, + 0.039164476096630096, + 0.01668434776365757, + 0.018286986276507378, + 0.028006810694932938, + 0.032958984375, + 0.02002651058137417, + 0.04073677584528923, + 0.03613337501883507, + 0.02841143123805523, + 0.016132960096001625, + 0.019453352317214012, + 0.019957372918725014, + 0.024772273376584053, + 0.017620911821722984, + 0.0262303426861763, + 0.016594287008047104, + 0.016468515619635582, + 0.00481930049136281, + 0.004819798748940229, + 0.0037694498896598816, + 0.004096970893442631, + 0.004478448070585728, + 0.0038508286233991385, + 0.0031192319002002478, + 0.004376793280243874, + 0.006870536133646965, + 0.010900729335844517, + 0.007277678698301315, + 0.01089709997177124, + 0.005137982778251171, + 0.011005370877683163, + 0.0051138694398105145, + 0.006569409742951393, + 0.004556629341095686, + 0.006657934281975031, + 0.004901152104139328, + 0.006799148861318827, + 0.004440720193088055, + 0.006136735901236534, + 0.004311367403715849, + 0.005086806137114763, + 0.005729632452130318, + 0.007134691346436739, + 0.009005633182823658, + 0.003871561260893941, + 0.00380415515974164, + 0.0059471712447702885, + 0.0037068563979119062, + 0.0046198125928640366, + 0.005447074770927429, + 0.00761836301535368, + 0.011554725468158722, + 0.005722825415432453, + 0.004275005776435137, + 0.007464459631592035, + 0.0034226924180984497, + 0.006819102913141251, + 0.005430083721876144, + 0.004777552094310522, + 0.005501365289092064, + 0.0033332575112581253, + 0.004339003004133701, + 0.011965920217335224, + 0.003968933597207069, + 0.014427701942622662, + 0.007921019569039345, + 0.007506001275032759, + 0.0070184762589633465, + 0.005805289838463068, + 0.008490748703479767, + 0.007911011576652527, + 0.006258921232074499, + 0.006947459653019905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0032165716402232647 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960546433925629 + ], + "opposite_logprob_tokens": [ + -5.130960464477539 + ], + "logprob_reward": -0.005960546433925629, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960546433925629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26261958479881287, + 0.09522101283073425, + 0.05961941182613373, + 0.07659537345170975, + 0.10734034329652786, + 0.048711903393268585, + 0.04472702741622925, + 0.04068048298358917, + 0.14309322834014893, + 0.12139160931110382 + ], + "credit_map": [ + 0.03980730101466179, + 0.02214411087334156, + 0.036333777010440826, + 0.023532649502158165, + 0.01586803048849106, + 0.04371935501694679, + 0.04073432832956314, + 0.04048002511262894, + 0.011577987112104893, + 0.008616854436695576, + 0.013818328268826008, + 0.009771529585123062, + 0.010330725461244583, + 0.01213089656084776, + 0.018374314531683922, + 0.010600387118756771, + 0.006755385547876358, + 0.008218909613788128, + 0.0069663734175264835, + 0.010844319127500057, + 0.005915751680731773, + 0.006897415034472942, + 0.007725676987320185, + 0.006295575760304928, + 0.011601547710597515, + 0.010732932947576046, + 0.00910245906561613, + 0.009801015257835388, + 0.009472041390836239, + 0.00839273352175951, + 0.007753552403301001, + 0.009739086031913757, + 0.011250471696257591, + 0.015175496228039265, + 0.012621882371604443, + 0.012721259146928787, + 0.012571927160024643, + 0.019279321655631065, + 0.01023401040583849, + 0.013485973700881004, + 0.00594807043671608, + 0.007483565714210272, + 0.007253824267536402, + 0.007640530820935965, + 0.004026117268949747, + 0.005874738097190857, + 0.004697934724390507, + 0.005787117872387171, + 0.005477712489664555, + 0.00547654926776886, + 0.005582416895776987, + 0.004659355618059635, + 0.004841959569603205, + 0.009857522323727608, + 0.0038169061299413443, + 0.005014607682824135, + 0.004704493563622236, + 0.005635239649564028, + 0.0048995730467140675, + 0.004371035844087601, + 0.0045529636554419994, + 0.006343386135995388, + 0.0035635698586702347, + 0.006610223092138767, + 0.028579553589224815, + 0.009397349320352077, + 0.017268311232328415, + 0.00708193751052022, + 0.007305488921701908, + 0.02304878830909729, + 0.009985114447772503, + 0.04042668640613556, + 0.014810444787144661, + 0.013012313283979893, + 0.015834828838706017, + 0.00925533939152956, + 0.01696356199681759, + 0.017153941094875336, + 0.013884393498301506, + 0.02047679014503956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.005960546433925629 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024873053189367056 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.0024873053189367056, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024873053189367056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17609083652496338, + 0.07086455076932907, + 0.11381421238183975, + 0.04785306751728058, + 0.05378037691116333, + 0.07694026827812195, + 0.08935260772705078, + 0.042234018445014954, + 0.28170081973075867, + 0.047369323670864105 + ], + "credit_map": [ + 0.02964371256530285, + 0.012311226688325405, + 0.03027057647705078, + 0.020476525649428368, + 0.008624326437711716, + 0.021328995004296303, + 0.030229344964027405, + 0.02320612035691738, + 0.009735282510519028, + 0.009896344505250454, + 0.010359032079577446, + 0.011911177076399326, + 0.0064321174286305904, + 0.007002927362918854, + 0.008006173186004162, + 0.007521491963416338, + 0.017131660133600235, + 0.012743487022817135, + 0.013532435521483421, + 0.02117452397942543, + 0.01059721875935793, + 0.009059719741344452, + 0.01362889539450407, + 0.015946269035339355, + 0.004542425740510225, + 0.007107274606823921, + 0.005596097093075514, + 0.006992178037762642, + 0.004856464918702841, + 0.004905954003334045, + 0.006937820464372635, + 0.006914852652698755, + 0.006634500343352556, + 0.006091505289077759, + 0.005410180892795324, + 0.01180433202534914, + 0.006080767139792442, + 0.007644070778042078, + 0.004393129609525204, + 0.005721887107938528, + 0.007905668579041958, + 0.012338230386376381, + 0.010809469036757946, + 0.009966720826923847, + 0.006668918766081333, + 0.011160697788000107, + 0.008209075778722763, + 0.009881485253572464, + 0.01130598783493042, + 0.014405165798962116, + 0.017855696380138397, + 0.015874193981289864, + 0.006227180827409029, + 0.00942270178347826, + 0.007912295870482922, + 0.006349373143166304, + 0.004172682762145996, + 0.006507329177111387, + 0.005885833874344826, + 0.00563401123508811, + 0.004928860813379288, + 0.0069619836285710335, + 0.004150270018726587, + 0.0039930460043251514, + 0.0960676521062851, + 0.025188108906149864, + 0.026309307664632797, + 0.01219644770026207, + 0.01912965066730976, + 0.06379950791597366, + 0.009685490280389786, + 0.029324626550078392, + 0.007090564351528883, + 0.005837634671479464, + 0.007896151393651962, + 0.005035074893385172, + 0.006064447574317455, + 0.007201286032795906, + 0.003748043207451701, + 0.00449611758813262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0024873053189367056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031571239233017 + ], + "opposite_logprob_tokens": [ + -2.350315809249878 + ], + "logprob_reward": -0.10031571239233017, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031571239233017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2514570951461792, + 0.08166511356830597, + 0.07232807576656342, + 0.0536290742456913, + 0.1571793556213379, + 0.07627422362565994, + 0.04822642356157303, + 0.05232328176498413, + 0.1618724912405014, + 0.045044928789138794 + ], + "credit_map": [ + 0.03551199287176132, + 0.022681739181280136, + 0.03165712580084801, + 0.02444799244403839, + 0.021161314100027084, + 0.042112596333026886, + 0.036491017788648605, + 0.03739330917596817, + 0.00918629765510559, + 0.007948176935315132, + 0.01186536904424429, + 0.008543032221496105, + 0.007168181706219912, + 0.011809847317636013, + 0.014250991865992546, + 0.010893207974731922, + 0.008413347415626049, + 0.008574290201067924, + 0.008026475086808205, + 0.013341362588107586, + 0.005539992358535528, + 0.008490522392094135, + 0.0106881745159626, + 0.00925391260534525, + 0.007329292595386505, + 0.0070116398856043816, + 0.007095403037965298, + 0.007622331380844116, + 0.005757829640060663, + 0.005533681716769934, + 0.005545320454984903, + 0.007733573205769062, + 0.014788851141929626, + 0.01905885711312294, + 0.017693428322672844, + 0.027751676738262177, + 0.018911562860012054, + 0.031305719166994095, + 0.010174083523452282, + 0.01749517396092415, + 0.010514209046959877, + 0.010489997453987598, + 0.010484647005796432, + 0.011514965444803238, + 0.006975503638386726, + 0.008636442013084888, + 0.007667293772101402, + 0.009991171769797802, + 0.006207963917404413, + 0.007373034488409758, + 0.009581577964127064, + 0.007664965000003576, + 0.004538916517049074, + 0.005696930456906557, + 0.0034036338329315186, + 0.0037594030145555735, + 0.006559302099049091, + 0.00789577979594469, + 0.006040896289050579, + 0.0047444929368793964, + 0.005479313898831606, + 0.006592526100575924, + 0.0048920996487140656, + 0.010118873789906502, + 0.04696646332740784, + 0.011901168152689934, + 0.017335807904601097, + 0.010063531808555126, + 0.011182085610926151, + 0.02357187680900097, + 0.009817910380661488, + 0.03103364072740078, + 0.004896891303360462, + 0.005209633614867926, + 0.004734820686280727, + 0.0048114764504134655, + 0.005384279880672693, + 0.0068750339560210705, + 0.00617642980068922, + 0.006956363096833229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.10031571239233017 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, captured from a level perspective, captures a moment steeped in a sense of heritage. Inside a dimly lit, antique-style Chinese medicine shop, an elderly master, his silver hair meticulously combed and his face lined with wisdom, wears a dark blue cotton and linen double-breasted jacket, sleeves slightly rolled up, revealing his veiny yet remarkably strong forearms. He calmly and solemnly hands over a carefully selected bundle of herbs. This bundle is truly extraordinary: a medley of earthy, dark brown roots, a few broad, unfurling dark green leaves, and a handful of small, lavender dried flowers, bound together with a simple hemp string. Across from him, a young apprentice bows slightly, hands respectfully extended, ready to receive. Dressed in a simple off-white, coarse cloth shirt, the apprentice's gaze is focused and reverent, his entire attention focused on his master's hands and the bundle. In the background stood a massive, neatly arranged medicine cabinet, its countless small drawers covered with yellowed paper labels. The air seemed to be filled with the faint aroma of medicine and dust. A ray of afternoon sunlight slanted in through the carved wooden window on the side, forming a sharp beam that perfectly illuminated the two people's joined hands and the bundle of herbs, gilding this tranquil, focused, and ritualistic scene with a warm and sacred golden aura.", + "seed": 21041, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_8/index_5621_seed_21041_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03271660655736923, + "vlm_logprob": -0.03271660655736923, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.03271660655736923, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030300721526145935 + ], + "opposite_logprob_tokens": [ + -3.9053006172180176 + ], + "logprob_reward": -0.030300721526145935, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.030300721526145935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29899701476097107, + 0.08698340505361557, + 0.21284979581832886, + 0.09800652414560318, + 0.05135606974363327, + 0.03208724036812782, + 0.042277686297893524, + 0.043452274054288864, + 0.05409432202577591, + 0.07989571988582611 + ], + "credit_map": [ + 0.021520832553505898, + 0.03521770238876343, + 0.033777277916669846, + 0.04420947656035423, + 0.036304980516433716, + 0.048854678869247437, + 0.04488038271665573, + 0.03423164412379265, + 0.007369714789092541, + 0.012862750329077244, + 0.008407206274569035, + 0.011642471887171268, + 0.013930498622357845, + 0.010367553681135178, + 0.012298968620598316, + 0.010104235261678696, + 0.013785053044557571, + 0.018750078976154327, + 0.012686614878475666, + 0.01624434068799019, + 0.01713227853178978, + 0.07470369338989258, + 0.01471069734543562, + 0.04483700916171074, + 0.005288494750857353, + 0.005702571477741003, + 0.005469938274472952, + 0.0048181465826928616, + 0.004241753835231066, + 0.035815030336380005, + 0.005944578908383846, + 0.030726000666618347, + 0.004231158643960953, + 0.011167913675308228, + 0.0033503547310829163, + 0.005186654161661863, + 0.005056718364357948, + 0.009034980088472366, + 0.006460066884756088, + 0.006868220400065184, + 0.0025724873412400484, + 0.005934051238000393, + 0.002290206728503108, + 0.004242626018822193, + 0.004272086080163717, + 0.004063419532030821, + 0.004579391796141863, + 0.004132966045290232, + 0.004287362564355135, + 0.005353351589292288, + 0.0035587791353464127, + 0.0034589653369039297, + 0.004240233916789293, + 0.010320735163986683, + 0.004952223971486092, + 0.006106032989919186, + 0.0036706821992993355, + 0.0036630842369049788, + 0.003487343667075038, + 0.002793816151097417, + 0.0032100523822009563, + 0.013161461800336838, + 0.002668880857527256, + 0.01079694926738739, + 0.004986709915101528, + 0.010656722821295261, + 0.004346999805420637, + 0.007827089168131351, + 0.006329867523163557, + 0.007228848058730364, + 0.007632600609213114, + 0.00508548179641366, + 0.005987732671201229, + 0.01687600091099739, + 0.006200592964887619, + 0.008911975659430027, + 0.007299656514078379, + 0.014995795674622059, + 0.008132776245474815, + 0.011491183191537857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.030300721526145935 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007649770472198725 + ], + "opposite_logprob_tokens": [ + -4.882649898529053 + ], + "logprob_reward": -0.007649770472198725, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007649770472198725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16500365734100342, + 0.14171041548252106, + 0.2651200592517853, + 0.08102649450302124, + 0.07265758514404297, + 0.03471965715289116, + 0.08259116113185883, + 0.03870109096169472, + 0.04245995730161667, + 0.07600988447666168 + ], + "credit_map": [ + 0.015251190401613712, + 0.031237857416272163, + 0.017723847180604935, + 0.019358478486537933, + 0.018660925328731537, + 0.025153804570436478, + 0.020813358947634697, + 0.016804208979010582, + 0.01343659870326519, + 0.028576910495758057, + 0.016956066712737083, + 0.019052714109420776, + 0.016806667670607567, + 0.0155820706859231, + 0.012618311680853367, + 0.01868109032511711, + 0.02458679862320423, + 0.021829118952155113, + 0.02376236766576767, + 0.02314693108201027, + 0.020315401256084442, + 0.09449305385351181, + 0.01708432286977768, + 0.03990209102630615, + 0.008339802734553814, + 0.022557038813829422, + 0.0074411844834685326, + 0.007725362665951252, + 0.007513967342674732, + 0.012550221756100655, + 0.007597888354212046, + 0.007301041390746832, + 0.0070387753657996655, + 0.020145868882536888, + 0.005708704236894846, + 0.007198627106845379, + 0.007462166249752045, + 0.010681066662073135, + 0.008164383471012115, + 0.006257995031774044, + 0.003322652308270335, + 0.005718853324651718, + 0.002927308203652501, + 0.004263032227754593, + 0.003286132588982582, + 0.004867845680564642, + 0.0033190702088177204, + 0.007014766801148653, + 0.006209590006619692, + 0.01224697008728981, + 0.005609069485217333, + 0.0047037177719175816, + 0.007536265533417463, + 0.024990249425172806, + 0.007489311974495649, + 0.013805992901325226, + 0.0023724588099867105, + 0.0043516154401004314, + 0.0024772973265498877, + 0.001912807929329574, + 0.0022742985747754574, + 0.013437749817967415, + 0.0024850736372172832, + 0.009389791637659073, + 0.003696904983371496, + 0.01055232621729374, + 0.0032831996213644743, + 0.0051199872978031635, + 0.004516147542744875, + 0.005319280084222555, + 0.005335856229066849, + 0.004636259749531746, + 0.004344794433563948, + 0.0022202946711331606, + 0.0044056689366698265, + 0.007226372137665749, + 0.002470838837325573, + 0.02865961380302906, + 0.0041154492646455765, + 0.02256685495376587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.007649770472198725 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595983536913991 + ], + "opposite_logprob_tokens": [ + -5.130959987640381 + ], + "logprob_reward": -0.00595983536913991, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595983536913991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23183982074260712, + 0.1277402639389038, + 0.3014161288738251, + 0.060344014316797256, + 0.036394152790308, + 0.02331525832414627, + 0.033662401139736176, + 0.10770002752542496, + 0.03796486556529999, + 0.039623066782951355 + ], + "credit_map": [ + 0.019424598664045334, + 0.030706778168678284, + 0.017180219292640686, + 0.034654177725315094, + 0.034759826958179474, + 0.03680720552802086, + 0.03297092765569687, + 0.02533608116209507, + 0.006195127032697201, + 0.021189427003264427, + 0.012478386983275414, + 0.02417108044028282, + 0.02326822839677334, + 0.014593126252293587, + 0.01668523997068405, + 0.009159637615084648, + 0.02153084985911846, + 0.019987652078270912, + 0.022659704089164734, + 0.023582616820931435, + 0.03674541413784027, + 0.09929526597261429, + 0.030989201739430428, + 0.046625424176454544, + 0.003499972401186824, + 0.0036812331527471542, + 0.004449640400707722, + 0.004554218612611294, + 0.005766559857875109, + 0.017889076843857765, + 0.004498832393437624, + 0.01600448228418827, + 0.002454377245157957, + 0.007420742884278297, + 0.0026058971416205168, + 0.004208623897284269, + 0.005548403598368168, + 0.005953625775873661, + 0.004244804847985506, + 0.003957677166908979, + 0.0011469037272036076, + 0.002092729089781642, + 0.0020183396991342306, + 0.0015621406491845846, + 0.0027854181826114655, + 0.0046785506419837475, + 0.002201198134571314, + 0.006829977035522461, + 0.0029319829773157835, + 0.006170893553644419, + 0.003016652539372444, + 0.004700545687228441, + 0.003058075439184904, + 0.006070565432310104, + 0.004068249836564064, + 0.0036454382352530956, + 0.006638252176344395, + 0.009671641513705254, + 0.005450955592095852, + 0.005412222817540169, + 0.005821918603032827, + 0.03741428628563881, + 0.007552923634648323, + 0.029737817123532295, + 0.0033778224606066942, + 0.006871315184980631, + 0.0027590077370405197, + 0.004346765577793121, + 0.004157227464020252, + 0.006516669876873493, + 0.005601625889539719, + 0.004334431607276201, + 0.003757381346076727, + 0.003389642806723714, + 0.0035297530703246593, + 0.0044668386690318584, + 0.002337440848350525, + 0.011259003542363644, + 0.004029919393360615, + 0.0068530854769051075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.00595983536913991 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732163019478321 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006732163019478321, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732163019478321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3310612738132477, + 0.07440806180238724, + 0.1742851287126541, + 0.10103554278612137, + 0.08456884324550629, + 0.023142145946621895, + 0.04438096284866333, + 0.04104685038328171, + 0.03195430338382721, + 0.09411687403917313 + ], + "credit_map": [ + 0.03542724996805191, + 0.06664920598268509, + 0.026890829205513, + 0.05889468267560005, + 0.02534494735300541, + 0.037351299077272415, + 0.0499018169939518, + 0.030601268634200096, + 0.006242020521312952, + 0.01241273619234562, + 0.008325573056936264, + 0.014058389700949192, + 0.00849604420363903, + 0.007948463782668114, + 0.007893676869571209, + 0.009031162597239017, + 0.016675129532814026, + 0.015033029951155186, + 0.015066252090036869, + 0.015328754670917988, + 0.009329385124146938, + 0.05700140446424484, + 0.011032238602638245, + 0.03481893241405487, + 0.006277880165725946, + 0.008725471794605255, + 0.008777115494012833, + 0.008099876344203949, + 0.011533403769135475, + 0.022022860124707222, + 0.015747981145977974, + 0.0198509618639946, + 0.004172769375145435, + 0.009918980300426483, + 0.004612623248249292, + 0.005821718834340572, + 0.007975521497428417, + 0.021460840478539467, + 0.009307311847805977, + 0.021299080923199654, + 0.0022195675410330296, + 0.002640026854351163, + 0.0028226240538060665, + 0.0044050500728189945, + 0.0023534544743597507, + 0.0023716141004115343, + 0.00385967711918056, + 0.002470135223120451, + 0.003636256791651249, + 0.009349208325147629, + 0.003767186775803566, + 0.005836006719619036, + 0.0033002905547618866, + 0.008455539122223854, + 0.004893083591014147, + 0.0051433914341032505, + 0.0025946232490241528, + 0.0028581060469150543, + 0.0027541627641767263, + 0.00279478682205081, + 0.002132267691195011, + 0.013851016759872437, + 0.0024608548264950514, + 0.0116010382771492, + 0.0028855730779469013, + 0.00828311126679182, + 0.002737312577664852, + 0.003782465821132064, + 0.0027603786438703537, + 0.004671963397413492, + 0.003188805188983679, + 0.003644695971161127, + 0.0064901621080935, + 0.0055342791602015495, + 0.006104123312979937, + 0.009624309837818146, + 0.003937209025025368, + 0.03241387754678726, + 0.007053582929074764, + 0.022959334775805473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.006732163019478321 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a dimly lit, antique-style Chinese medicine shop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly master with silver hair wearing a dark blue double-breasted jacket visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bundle of herbs containing dark brown roots, dark green leaves, and lavender dried flowers bound with hemp string visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a young apprentice in an off-white shirt bowing slightly with hands extended to receive the herbs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a ray of afternoon sunlight forming a sharp beam that illuminates the joined hands and the bundle of herbs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294054239988327 + ], + "opposite_logprob_tokens": [ + -2.237940549850464 + ], + "logprob_reward": -0.11294054239988327, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294054239988327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38277944922447205, + 0.0977475643157959, + 0.08984404057264328, + 0.07502347230911255, + 0.057130102068185806, + 0.04384680464863777, + 0.08869502693414688, + 0.05686849355697632, + 0.03815585747361183, + 0.06990919262170792 + ], + "credit_map": [ + 0.041377563029527664, + 0.08287772536277771, + 0.026925992220640182, + 0.06644512712955475, + 0.03824874758720398, + 0.04553532972931862, + 0.05213205888867378, + 0.02923690341413021, + 0.005893518682569265, + 0.012690330855548382, + 0.008859526365995407, + 0.022325817495584488, + 0.014427457936108112, + 0.01268862746655941, + 0.010357815772294998, + 0.010504467412829399, + 0.00765026081353426, + 0.007902148179709911, + 0.008101768791675568, + 0.010445755906403065, + 0.007511827163398266, + 0.024742955341935158, + 0.012176238000392914, + 0.011313088238239288, + 0.004830651450902224, + 0.009023449383676052, + 0.004049327690154314, + 0.00609339214861393, + 0.00760505348443985, + 0.015663689002394676, + 0.013030687347054482, + 0.014727221801877022, + 0.0035054259933531284, + 0.007065574172884226, + 0.003456729231402278, + 0.005355627741664648, + 0.007981183007359505, + 0.01352124847471714, + 0.006017718464136124, + 0.010226600803434849, + 0.004040009342133999, + 0.004633207805454731, + 0.0037812809459865093, + 0.004816255997866392, + 0.007338560651987791, + 0.008279804140329361, + 0.004964314866811037, + 0.005993372295051813, + 0.005632943939417601, + 0.00856272503733635, + 0.004623552784323692, + 0.006613892503082752, + 0.013051997870206833, + 0.023993192240595818, + 0.008781441487371922, + 0.017435282468795776, + 0.0036492375656962395, + 0.0033838460221886635, + 0.0031963842920958996, + 0.003602284472435713, + 0.0033540299627929926, + 0.018497418612241745, + 0.004279986955225468, + 0.01690530590713024, + 0.003064685733988881, + 0.01004057563841343, + 0.0035904168616980314, + 0.004419428762048483, + 0.0037136913742870092, + 0.0042648883536458015, + 0.0050588552840054035, + 0.004003313370049, + 0.005149457603693008, + 0.007709137629717588, + 0.004937727935612202, + 0.007836815901100636, + 0.004353875294327736, + 0.020089834928512573, + 0.006283500697463751, + 0.013548847287893295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.11294054239988327 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_9/index_tensor([1096, 1096])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_9/index_tensor([1096, 1096])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5535dcbb9448f65c6cc857d774590687d09f85c3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_9/index_tensor([1096, 1096])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly low-angle close-up transports the viewer into a surreal digital rendering of a dense, dwarf bamboo forest, set against a dark, invisible canopy. The bamboo's unusually broad, waxy leaves are a deep green, their surface gleaming with a damp sheen in the light, while their slender, segmented stems are sturdy. The entire scene is dramatically illuminated by a powerful beam of sunlight, cutting diagonally through the darkness. This light source is not constant; it flickers violently, casting the environment in rapid, stroboscopic pulses, alternating between dazzling light and deep shadows. With each golden flash, the bamboo is captured in a subtle, time-lapse-like movement, its leaves and stems delicately twisting and bending, intermittently toward the beam. In the shimmering air, tiny dust particles dance erratically, appearing and disappearing in the stroboscopic shafts of light, while shimmering dewdrops on the leaves catch each pulse, creating momentary, brilliant crystal flickers that shimmer with prismatic intensity, generating a captivating and intense atmosphere of energy.", + "seed": 1317, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_9/index_1096_seed_1317_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.24689242504537107, + "vlm_logprob": -0.24689242504537107, + "vlm_yes_no_margin": 2.9750000119209288, + "vlm_token_credit": -0.24689242504537107, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11688639223575592 + ], + "opposite_logprob_tokens": [ + -2.2418863773345947 + ], + "logprob_reward": -0.11688639223575592, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11688639223575592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08846697211265564, + 0.11320748925209045, + 0.0973244458436966, + 0.24963098764419556, + 0.08728217333555222, + 0.05162378028035164, + 0.035409342497587204, + 0.05766432359814644, + 0.10523917526006699, + 0.11415138840675354 + ], + "credit_map": [ + 0.005958497989922762, + 0.006496286019682884, + 0.008611016906797886, + 0.009097293019294739, + 0.006253259256482124, + 0.021810224279761314, + 0.01412924937903881, + 0.016111133620142937, + 0.006576063111424446, + 0.00840380322188139, + 0.010503588244318962, + 0.009011482819914818, + 0.010097058489918709, + 0.020352214574813843, + 0.01993185468018055, + 0.02833140827715397, + 0.011515644378960133, + 0.008224444463849068, + 0.009303365834057331, + 0.006810975726693869, + 0.010477552190423012, + 0.015183779411017895, + 0.0127033032476902, + 0.02310536801815033, + 0.015448675490915775, + 0.011570572853088379, + 0.016889363527297974, + 0.016035497188568115, + 0.019230540841817856, + 0.06063039228320122, + 0.04049916937947273, + 0.06932675838470459, + 0.00589381018653512, + 0.01038536336272955, + 0.010276705026626587, + 0.009661211632192135, + 0.010899602435529232, + 0.01849309168756008, + 0.006127998698502779, + 0.015544377267360687, + 0.004076462239027023, + 0.0041113924235105515, + 0.012601243332028389, + 0.0032266464550048113, + 0.00525608379393816, + 0.013215838931500912, + 0.003709311131387949, + 0.00542679475620389, + 0.0035816270392388105, + 0.002572833327576518, + 0.008610138669610023, + 0.0032490077428519726, + 0.002985255792737007, + 0.0034094490110874176, + 0.003837668802589178, + 0.007163361180573702, + 0.004856035113334656, + 0.004328618291765451, + 0.009796933270990849, + 0.004800553899258375, + 0.003625214332714677, + 0.01622488908469677, + 0.004983298480510712, + 0.00904877483844757, + 0.007855293340981007, + 0.007986290380358696, + 0.020234240218997, + 0.008174451068043709, + 0.008502122014760971, + 0.015561744570732117, + 0.010156202130019665, + 0.026768825948238373, + 0.009149227291345596, + 0.009950713254511356, + 0.013626271858811378, + 0.014751897193491459, + 0.01239231787621975, + 0.012917057611048222, + 0.026921983808279037, + 0.014441902749240398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.11688639223575592 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556896544992924 + ], + "opposite_logprob_tokens": [ + -4.387557029724121 + ], + "logprob_reward": -0.012556896544992924, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556896544992924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15201568603515625, + 0.07543237507343292, + 0.032615043222904205, + 0.09779572486877441, + 0.03607332333922386, + 0.10778649896383286, + 0.2853715419769287, + 0.04517076537013054, + 0.07575798779726028, + 0.09198108315467834 + ], + "credit_map": [ + 0.007615944836288691, + 0.006685562431812286, + 0.009526384063065052, + 0.008299026638269424, + 0.007671296130865812, + 0.051603108644485474, + 0.01970708556473255, + 0.040907278656959534, + 0.004703504499047995, + 0.005501460283994675, + 0.011129830963909626, + 0.006693401373922825, + 0.006134419236332178, + 0.012423415668308735, + 0.01199706643819809, + 0.016849277541041374, + 0.0028891321271657944, + 0.0026734364219009876, + 0.004036670550704002, + 0.00329976761713624, + 0.0036680910270661116, + 0.0060326699167490005, + 0.003723045578226447, + 0.006292230915278196, + 0.004078669007867575, + 0.0033107276540249586, + 0.006052335258573294, + 0.0028496647719293833, + 0.005896031390875578, + 0.03370845690369606, + 0.005528321955353022, + 0.036371514201164246, + 0.003092107828706503, + 0.002787406090646982, + 0.003655193606391549, + 0.005101384129375219, + 0.0038834726437926292, + 0.007278340868651867, + 0.0037714047357439995, + 0.006504013668745756, + 0.007362499833106995, + 0.005555783864110708, + 0.0250582005828619, + 0.006356330588459969, + 0.009705481119453907, + 0.03545915707945824, + 0.006766655948013067, + 0.011522384360432625, + 0.023772424086928368, + 0.012564927339553833, + 0.058705173432826996, + 0.010750354267656803, + 0.01746504008769989, + 0.10365649312734604, + 0.013303399085998535, + 0.045153725892305374, + 0.005185404792428017, + 0.0036115157417953014, + 0.00966574065387249, + 0.0027717093471437693, + 0.002812327817082405, + 0.0096941739320755, + 0.003413788741454482, + 0.008016103878617287, + 0.0063437726348638535, + 0.00526466453447938, + 0.015297573991119862, + 0.006869128905236721, + 0.006722441408783197, + 0.013691197149455547, + 0.005091725382953882, + 0.016477486118674278, + 0.008650103583931923, + 0.009206688962876797, + 0.011843535117805004, + 0.010207978077232838, + 0.009779481217265129, + 0.009165181778371334, + 0.021900638937950134, + 0.01122747827321291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.012556896544992924 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002491348423063755 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.002491348423063755, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002491348423063755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4025368094444275, + 0.07433607429265976, + 0.08895468711853027, + 0.04246572032570839, + 0.04041384533047676, + 0.08463414013385773, + 0.08912093192338943, + 0.11114273220300674, + 0.030547108501195908, + 0.03584787994623184 + ], + "credit_map": [ + 0.013820902444422245, + 0.011028766632080078, + 0.01667349599301815, + 0.012788940221071243, + 0.01714089699089527, + 0.15750519931316376, + 0.02711418829858303, + 0.14646446704864502, + 0.0052563645876944065, + 0.005167874041944742, + 0.008145622909069061, + 0.007352777291089296, + 0.0075329868122935295, + 0.016441354528069496, + 0.009276945143938065, + 0.015162164345383644, + 0.009480638429522514, + 0.00865546241402626, + 0.009014440700411797, + 0.007662292569875717, + 0.009827897883951664, + 0.01815107651054859, + 0.007819878868758678, + 0.018343007192015648, + 0.004450679756700993, + 0.003284245962277055, + 0.005258344579488039, + 0.0037781985010951757, + 0.0050961789675056934, + 0.005888316780328751, + 0.009663814678788185, + 0.005045944359153509, + 0.003319954499602318, + 0.004023149609565735, + 0.005646958947181702, + 0.004968785680830479, + 0.004417910240590572, + 0.009213116951286793, + 0.002813557628542185, + 0.00601041316986084, + 0.008326482027769089, + 0.00553005700930953, + 0.021830249577760696, + 0.005625370889902115, + 0.012111612595617771, + 0.009479457512497902, + 0.006428959779441357, + 0.015301954932510853, + 0.009776817634701729, + 0.00962016824632883, + 0.019685452803969383, + 0.00832301564514637, + 0.009335929527878761, + 0.006478044670075178, + 0.008153935894370079, + 0.017747579142451286, + 0.008676044642925262, + 0.01026934664696455, + 0.01695605181157589, + 0.010694046504795551, + 0.00660815229639411, + 0.02516132965683937, + 0.010633880272507668, + 0.0221438966691494, + 0.0023116001393646, + 0.003526396583765745, + 0.005282765720039606, + 0.0030767559073865414, + 0.002097390592098236, + 0.004923857748508453, + 0.0036388286389410496, + 0.005689514335244894, + 0.0031651973258703947, + 0.0034068552777171135, + 0.004828849341720343, + 0.003558629658073187, + 0.0035737704019993544, + 0.0040434845723211765, + 0.007264341227710247, + 0.006006754469126463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.002491348423063755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.053865671157837 + ], + "opposite_logprob_tokens": [ + -0.4288657307624817 + ], + "logprob_reward": -1.053865671157837, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.053865671157837, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10731038451194763, + 0.13986395299434662, + 0.10109294205904007, + 0.11452681571245193, + 0.09338437765836716, + 0.0961494967341423, + 0.11029403656721115, + 0.1111283227801323, + 0.03372115641832352, + 0.0925285592675209 + ], + "credit_map": [ + 0.003906445112079382, + 0.00549422949552536, + 0.006485233549028635, + 0.006525612436234951, + 0.010886146686971188, + 0.030544331297278404, + 0.0173957422375679, + 0.026072639971971512, + 0.011287091299891472, + 0.012067626230418682, + 0.013381223194301128, + 0.01020019967108965, + 0.009771713986992836, + 0.027398603036999702, + 0.020588018000125885, + 0.03516946732997894, + 0.013018423691391945, + 0.007149889599531889, + 0.014453732408583164, + 0.007666891440749168, + 0.008433610200881958, + 0.026948435232043266, + 0.00659991055727005, + 0.016822047531604767, + 0.004754941910505295, + 0.005948403384536505, + 0.009358658455312252, + 0.005807077512145042, + 0.008074854500591755, + 0.031184013932943344, + 0.014911861158907413, + 0.034487005323171616, + 0.00759558379650116, + 0.01105879433453083, + 0.010940272361040115, + 0.013281594030559063, + 0.010661762207746506, + 0.01993556134402752, + 0.007731224875897169, + 0.012179584242403507, + 0.006268167868256569, + 0.009627324528992176, + 0.010267907753586769, + 0.008641975000500679, + 0.014582967385649681, + 0.015847651287913322, + 0.0106352549046278, + 0.02027825266122818, + 0.010679731145501137, + 0.007935654371976852, + 0.0271309781819582, + 0.008494779467582703, + 0.00784915592521429, + 0.0234442800283432, + 0.006612071767449379, + 0.018147381022572517, + 0.008675671182572842, + 0.008381634950637817, + 0.019389789551496506, + 0.012134519405663013, + 0.006269311532378197, + 0.02554023079574108, + 0.007901689037680626, + 0.022835472598671913, + 0.0035284056793898344, + 0.002817779080942273, + 0.008213590830564499, + 0.004009487573057413, + 0.002146064769476652, + 0.004382974933832884, + 0.002653078641742468, + 0.005969780031591654, + 0.006404953543096781, + 0.013267183676362038, + 0.01714860275387764, + 0.007389744743704796, + 0.012865063734352589, + 0.010828748345375061, + 0.008032750338315964, + 0.01659150794148445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -1.053865671157837 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048661816865205765 + ], + "opposite_logprob_tokens": [ + -3.048661708831787 + ], + "logprob_reward": -0.048661816865205765, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048661816865205765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1385485976934433, + 0.07723653316497803, + 0.060592520982027054, + 0.10103055089712143, + 0.07444357126951218, + 0.12874095141887665, + 0.27192264795303345, + 0.0783601850271225, + 0.024708129465579987, + 0.04441632702946663 + ], + "credit_map": [ + 0.004739655647426844, + 0.00549285626038909, + 0.009232466109097004, + 0.004628583788871765, + 0.006833476480096579, + 0.048463430255651474, + 0.011194396764039993, + 0.047963742166757584, + 0.004598828963935375, + 0.006333569996058941, + 0.008213529363274574, + 0.008088555186986923, + 0.008935547433793545, + 0.01234059501439333, + 0.012036196887493134, + 0.016689708456397057, + 0.00861749891191721, + 0.005817164666950703, + 0.006552345119416714, + 0.004616245627403259, + 0.005348891951143742, + 0.0115730669349432, + 0.0060330647975206375, + 0.012034241110086441, + 0.004465299658477306, + 0.004426879808306694, + 0.006969938054680824, + 0.005736046005040407, + 0.0058778454549610615, + 0.02758866921067238, + 0.012986495159566402, + 0.03297937661409378, + 0.006269303150475025, + 0.008826196193695068, + 0.008580897934734821, + 0.0086964787915349, + 0.0088136475533247, + 0.016429800540208817, + 0.005394902545958757, + 0.01143234595656395, + 0.007825191132724285, + 0.008650005795061588, + 0.031455811113119125, + 0.009564780630171299, + 0.01282904576510191, + 0.03281853720545769, + 0.012175788171589375, + 0.013421790674328804, + 0.021478697657585144, + 0.010682307183742523, + 0.055750373750925064, + 0.010354950092732906, + 0.01591533049941063, + 0.10160451382398605, + 0.011031248606741428, + 0.04510520398616791, + 0.006967856083065271, + 0.006370642688125372, + 0.01387713197618723, + 0.01026691123843193, + 0.005686729680746794, + 0.01815737970173359, + 0.005895526148378849, + 0.011138007044792175, + 0.0024021959397941828, + 0.0017669890075922012, + 0.004635936580598354, + 0.0019508353434503078, + 0.002666934859007597, + 0.0034641562961041927, + 0.002241455251350999, + 0.005579627119004726, + 0.003316316520795226, + 0.006234719417989254, + 0.00713355140760541, + 0.004109919536858797, + 0.00734173646196723, + 0.005782983731478453, + 0.003345631994307041, + 0.007151471450924873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.048661816865205765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up transports the viewer into a surreal digital rendering of a dense, dwarf bamboo forest, set against a dark, invisible canopy. The bamboo's unusually broad, waxy leaves are a deep green, their surface gleaming with a damp sheen in the light, while their slender, segmented stems are sturdy. The entire scene is dramatically illuminated by a powerful beam of sunlight, cutting diagonally through the darkness. This light source is not constant; it flickers violently, casting the environment in rapid, stroboscopic pulses, alternating between dazzling light and deep shadows. With each golden flash, the bamboo is captured in a subtle, time-lapse-like movement, its leaves and stems delicately twisting and bending, intermittently toward the beam. In the shimmering air, tiny dust particles dance erratically, appearing and disappearing in the stroboscopic shafts of light, while shimmering dewdrops on the leaves catch each pulse, creating momentary, brilliant crystal flickers that shimmer with prismatic intensity, generating a captivating and intense atmosphere of energy.", + "seed": 1318, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_9/index_1096_seed_1318_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15287372027523816, + "vlm_logprob": -0.15287372027523816, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.15287372027523816, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17388933897018433 + ], + "opposite_logprob_tokens": [ + -1.923889398574829 + ], + "logprob_reward": -0.17388933897018433, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17388933897018433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12217244505882263, + 0.040591441094875336, + 0.044283676892519, + 0.046593669801950455, + 0.08983967453241348, + 0.06134657561779022, + 0.22875584661960602, + 0.09741753339767456, + 0.19843098521232605, + 0.07056820392608643 + ], + "credit_map": [ + 0.013407560996711254, + 0.02749103307723999, + 0.013519607484340668, + 0.01497614849358797, + 0.012042371556162834, + 0.017919160425662994, + 0.01149838138371706, + 0.011318182572722435, + 0.004979003220796585, + 0.006066254340112209, + 0.0057038734667003155, + 0.004723684396594763, + 0.004004948772490025, + 0.006372668780386448, + 0.003982985857874155, + 0.004758020862936974, + 0.004703683312982321, + 0.008398600853979588, + 0.007003836799412966, + 0.005319744814187288, + 0.0036029047332704067, + 0.007693804334849119, + 0.0034666501451283693, + 0.004094452131539583, + 0.004230401013046503, + 0.006153779104351997, + 0.00912682805210352, + 0.0055216774344444275, + 0.004850344732403755, + 0.00699969008564949, + 0.004512073006480932, + 0.005198877304792404, + 0.007581422105431557, + 0.009894931688904762, + 0.012991949915885925, + 0.014985752291977406, + 0.009517618454992771, + 0.016864914447069168, + 0.009005118161439896, + 0.008997957222163677, + 0.005183964502066374, + 0.0047721765004098415, + 0.013965478166937828, + 0.007683152798563242, + 0.0029396647587418556, + 0.009252590127289295, + 0.0032996118534356356, + 0.014249935746192932, + 0.033940982073545456, + 0.026893392205238342, + 0.039045300334692, + 0.03221551328897476, + 0.0202060267329216, + 0.03098621591925621, + 0.01959175057709217, + 0.02587665244936943, + 0.005376385059207678, + 0.00999443605542183, + 0.00509050814434886, + 0.00618715351447463, + 0.004246297292411327, + 0.008256283588707447, + 0.03086107224225998, + 0.027405399829149246, + 0.014954950660467148, + 0.012097416445612907, + 0.05606423690915108, + 0.026939719915390015, + 0.010105825029313564, + 0.03288138657808304, + 0.020916834473609924, + 0.024470612406730652, + 0.005082508083432913, + 0.00664126081392169, + 0.007387992460280657, + 0.006581844296306372, + 0.0034746313467621803, + 0.022106250748038292, + 0.007867020554840565, + 0.011426696553826332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.17388933897018433 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337296798825264 + ], + "opposite_logprob_tokens": [ + -3.4087297916412354 + ], + "logprob_reward": -0.0337296798825264, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337296798825264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07508604228496552, + 0.05615174025297165, + 0.047519464045763016, + 0.043142206966876984, + 0.05150645971298218, + 0.06306149810552597, + 0.5345168709754944, + 0.03712349385023117, + 0.046013008803129196, + 0.04587922990322113 + ], + "credit_map": [ + 0.008176509290933609, + 0.012747888453304768, + 0.012263170443475246, + 0.008986961096525192, + 0.006917773280292749, + 0.01025080494582653, + 0.007794197648763657, + 0.007948732003569603, + 0.007062852382659912, + 0.00899383146315813, + 0.007338536903262138, + 0.009196375496685505, + 0.006074049044400454, + 0.006990100722759962, + 0.005429230630397797, + 0.005066761747002602, + 0.0034405982587486506, + 0.0054358323104679585, + 0.009006760083138943, + 0.004061384592205286, + 0.003907239064574242, + 0.012985305860638618, + 0.002907790709286928, + 0.005774547811597586, + 0.0034355020616203547, + 0.0050545986741781235, + 0.007896333001554012, + 0.004253126215189695, + 0.004249434918165207, + 0.00888704415410757, + 0.003975293133407831, + 0.005390869919210672, + 0.004709547385573387, + 0.005645034834742546, + 0.007474064361304045, + 0.009433110244572163, + 0.005289069842547178, + 0.008948840200901031, + 0.0057411217130720615, + 0.004265667404979467, + 0.004638759884983301, + 0.0046936203725636005, + 0.012830953113734722, + 0.008915483020246029, + 0.004166759550571442, + 0.009690185077488422, + 0.006307126954197884, + 0.011818604543805122, + 0.09666550904512405, + 0.06037966161966324, + 0.10942749679088593, + 0.08198042213916779, + 0.026109375059604645, + 0.053052160888910294, + 0.044137608259916306, + 0.06276465207338333, + 0.0018246673280373216, + 0.0021871409844607115, + 0.003691874211654067, + 0.0026446182746440172, + 0.0015123669290915132, + 0.003840844612568617, + 0.01082509197294712, + 0.01059688813984394, + 0.003325172234326601, + 0.0026936603244394064, + 0.01465737447142601, + 0.004752443637698889, + 0.002235202817246318, + 0.007775777485221624, + 0.0051057348027825356, + 0.0054676346480846405, + 0.004753494635224342, + 0.005429648328572512, + 0.0065627568401396275, + 0.0049322969280183315, + 0.004952269606292248, + 0.008640853688120842, + 0.005783692933619022, + 0.004824213683605194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0337296798825264 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036220441106706858 + ], + "opposite_logprob_tokens": [ + -5.628622055053711 + ], + "logprob_reward": -0.0036220441106706858, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036220441106706858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09867949038743973, + 0.08135993033647537, + 0.028795436024665833, + 0.039680637419223785, + 0.1003432497382164, + 0.04364752396941185, + 0.23048533499240875, + 0.17736048996448517, + 0.10827502608299255, + 0.09137285500764847 + ], + "credit_map": [ + 0.011581644415855408, + 0.01954800821840763, + 0.011998488567769527, + 0.011547294445335865, + 0.012208964675664902, + 0.011459825560450554, + 0.012155154719948769, + 0.008180114440619946, + 0.011203553527593613, + 0.008446400053799152, + 0.013687130063772202, + 0.007869930937886238, + 0.007201052736490965, + 0.014262245036661625, + 0.009304046630859375, + 0.009385565295815468, + 0.002760829869657755, + 0.004404168110340834, + 0.00453721871599555, + 0.003484798828139901, + 0.003008896019309759, + 0.005167377181351185, + 0.0024532703682780266, + 0.002978879027068615, + 0.003816627897322178, + 0.00512453680858016, + 0.007171455305069685, + 0.006084425840526819, + 0.0034469575621187687, + 0.005884613376110792, + 0.0034227906726300716, + 0.0047292341478168964, + 0.009351431392133236, + 0.009859642945230007, + 0.015321740880608559, + 0.01347610354423523, + 0.011213806457817554, + 0.020577894523739815, + 0.010695644654333591, + 0.009846982546150684, + 0.0028761010617017746, + 0.002912995172664523, + 0.007297039497643709, + 0.003560808254405856, + 0.002830484649166465, + 0.012408013455569744, + 0.003977866377681494, + 0.007784216199070215, + 0.01990276202559471, + 0.02281379885971546, + 0.029684696346521378, + 0.027358820661902428, + 0.022734899073839188, + 0.03334948420524597, + 0.031799934804439545, + 0.04284093156456947, + 0.007502177730202675, + 0.00794583186507225, + 0.009048408828675747, + 0.011254862882196903, + 0.008179116062819958, + 0.02138546109199524, + 0.056951627135276794, + 0.05509301275014877, + 0.007273703347891569, + 0.0054937549866735935, + 0.03223395347595215, + 0.01299781259149313, + 0.0048743863590061665, + 0.0215662382543087, + 0.0099021652713418, + 0.013933011330664158, + 0.006080990191549063, + 0.007742487359791994, + 0.010926616378128529, + 0.006262612529098988, + 0.010430359281599522, + 0.027886902913451195, + 0.010574289597570896, + 0.01146859209984541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.0036220441106706858 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233258605003357 + ], + "opposite_logprob_tokens": [ + -0.8983258605003357 + ], + "logprob_reward": -0.5233258605003357, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233258605003357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08773618936538696, + 0.04205654188990593, + 0.03438536077737808, + 0.047792188823223114, + 0.06410128623247147, + 0.06269901245832443, + 0.206728994846344, + 0.07285884022712708, + 0.1970045119524002, + 0.18463708460330963 + ], + "credit_map": [ + 0.010843072086572647, + 0.01932481676340103, + 0.010652653872966766, + 0.01032405998557806, + 0.00769665464758873, + 0.010589426383376122, + 0.008221774362027645, + 0.010083731263875961, + 0.0048034414649009705, + 0.005718124099075794, + 0.007064393255859613, + 0.005725004244595766, + 0.0039263698272407055, + 0.00585861736908555, + 0.004116427153348923, + 0.004844162613153458, + 0.0043001919984817505, + 0.005079895723611116, + 0.006215142086148262, + 0.003808247856795788, + 0.0024993896950036287, + 0.005863472819328308, + 0.002594664925709367, + 0.004024353809654713, + 0.004429667256772518, + 0.005775269120931625, + 0.0075603979639709, + 0.006631033029407263, + 0.005776203237473965, + 0.008097886107861996, + 0.005276985466480255, + 0.004244746640324593, + 0.004349981434643269, + 0.007713569328188896, + 0.01061677373945713, + 0.007356118876487017, + 0.005805314984172583, + 0.009950605221092701, + 0.011772248893976212, + 0.006536670494824648, + 0.00441065663471818, + 0.004742552526295185, + 0.013134526088833809, + 0.00853007659316063, + 0.0033814532216638327, + 0.011420367285609245, + 0.005128819495439529, + 0.01195056363940239, + 0.03112749382853508, + 0.027716215699911118, + 0.031425777822732925, + 0.03054073080420494, + 0.019992586225271225, + 0.022391701117157936, + 0.017958302050828934, + 0.025576187297701836, + 0.003996941260993481, + 0.004396088887006044, + 0.005019382573664188, + 0.0061198449693620205, + 0.005591207649558783, + 0.012352175079286098, + 0.016948312520980835, + 0.018434889614582062, + 0.01877651736140251, + 0.023709852248430252, + 0.046314042061567307, + 0.02784411609172821, + 0.01136684324592352, + 0.028763212263584137, + 0.01886851340532303, + 0.021361416205763817, + 0.011637557297945023, + 0.016123756766319275, + 0.018291782587766647, + 0.015104171819984913, + 0.011369766667485237, + 0.059747364372015, + 0.023945802822709084, + 0.028416886925697327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.5233258605003357 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980167791247368 + ], + "opposite_logprob_tokens": [ + -3.529801607131958 + ], + "logprob_reward": -0.02980167791247368, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980167791247368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10826173424720764, + 0.056151509284973145, + 0.0747905746102333, + 0.18179106712341309, + 0.05292961746454239, + 0.04096851497888565, + 0.1952134668827057, + 0.06495869904756546, + 0.1152782216668129, + 0.10965661704540253 + ], + "credit_map": [ + 0.016053618863224983, + 0.011595476418733597, + 0.019006110727787018, + 0.013465448282659054, + 0.008918032050132751, + 0.016609657555818558, + 0.00843571126461029, + 0.01417767908424139, + 0.005512984003871679, + 0.00818636268377304, + 0.00835355743765831, + 0.006826519966125488, + 0.00869423896074295, + 0.007081051357090473, + 0.006317985709756613, + 0.005178809631615877, + 0.006141715217381716, + 0.00985496211796999, + 0.01256784051656723, + 0.008716979995369911, + 0.008553926832973957, + 0.01311987079679966, + 0.008277659304440022, + 0.007557614706456661, + 0.01916319504380226, + 0.027108658105134964, + 0.030893610790371895, + 0.029302392154932022, + 0.019201381132006645, + 0.0217796228826046, + 0.017810292541980743, + 0.016531921923160553, + 0.0045854151248931885, + 0.0058001079596579075, + 0.008614818565547466, + 0.006780985742807388, + 0.004531575832515955, + 0.011946775950491428, + 0.005373718217015266, + 0.005296217277646065, + 0.0031347398180514574, + 0.0040458794683218, + 0.005644817370921373, + 0.00742421206086874, + 0.0043030292727053165, + 0.0065677594393491745, + 0.003175393445417285, + 0.00667268643155694, + 0.03149521350860596, + 0.017238682135939598, + 0.04546857625246048, + 0.028563769534230232, + 0.012855559587478638, + 0.022639015689492226, + 0.013969136402010918, + 0.022983500733971596, + 0.003776852274313569, + 0.00562059273943305, + 0.004886250011622906, + 0.006276221945881844, + 0.00403094757348299, + 0.007448556367307901, + 0.016755014657974243, + 0.01616426557302475, + 0.015178435482084751, + 0.02397051826119423, + 0.013429849408566952, + 0.014564678072929382, + 0.007787120062857866, + 0.020466072484850883, + 0.009954677894711494, + 0.009926865808665752, + 0.008473384194076061, + 0.007936745882034302, + 0.01741316169500351, + 0.008186280727386475, + 0.007110337261110544, + 0.03374283388257027, + 0.013043833896517754, + 0.013750034384429455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.02980167791247368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up transports the viewer into a surreal digital rendering of a dense, dwarf bamboo forest, set against a dark, invisible canopy. The bamboo's unusually broad, waxy leaves are a deep green, their surface gleaming with a damp sheen in the light, while their slender, segmented stems are sturdy. The entire scene is dramatically illuminated by a powerful beam of sunlight, cutting diagonally through the darkness. This light source is not constant; it flickers violently, casting the environment in rapid, stroboscopic pulses, alternating between dazzling light and deep shadows. With each golden flash, the bamboo is captured in a subtle, time-lapse-like movement, its leaves and stems delicately twisting and bending, intermittently toward the beam. In the shimmering air, tiny dust particles dance erratically, appearing and disappearing in the stroboscopic shafts of light, while shimmering dewdrops on the leaves catch each pulse, creating momentary, brilliant crystal flickers that shimmer with prismatic intensity, generating a captivating and intense atmosphere of energy.", + "seed": 1319, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_9/index_1096_seed_1319_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1488971386104822, + "vlm_logprob": -0.1488971386104822, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.1488971386104822, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09981244802474976 + ], + "opposite_logprob_tokens": [ + -2.4748125076293945 + ], + "logprob_reward": -0.09981244802474976, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09981244802474976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30565229058265686, + 0.12530529499053955, + 0.06321161240339279, + 0.15687508881092072, + 0.09749071300029755, + 0.04257436469197273, + 0.06318167597055435, + 0.02600172534584999, + 0.05306281894445419, + 0.06664443761110306 + ], + "credit_map": [ + 0.030226804316043854, + 0.03291857987642288, + 0.053945157676935196, + 0.07343768328428268, + 0.0258690994232893, + 0.026379389688372612, + 0.031460244208574295, + 0.031415317207574844, + 0.009702719748020172, + 0.018720576539635658, + 0.018054038286209106, + 0.02385062351822853, + 0.011106129735708237, + 0.009736482053995132, + 0.012021681293845177, + 0.022113032639026642, + 0.004959069658070803, + 0.009258485399186611, + 0.008777353912591934, + 0.009135592728853226, + 0.008758720010519028, + 0.006672537885606289, + 0.008247319608926773, + 0.007402525283396244, + 0.013829619623720646, + 0.024100303649902344, + 0.025607967749238014, + 0.027034355327486992, + 0.02123529091477394, + 0.013631264679133892, + 0.01727321930229664, + 0.014163058251142502, + 0.01552351750433445, + 0.014479158446192741, + 0.014535450376570225, + 0.0165120679885149, + 0.010601356625556946, + 0.00821354053914547, + 0.009136631153523922, + 0.00848898570984602, + 0.003805336309596896, + 0.006253390107303858, + 0.00669870525598526, + 0.008465318009257317, + 0.00421407213434577, + 0.0034983125515282154, + 0.005942410323768854, + 0.003696815110743046, + 0.009339745156466961, + 0.008171167224645615, + 0.01104540191590786, + 0.014859167858958244, + 0.0035934012848883867, + 0.0038150132168084383, + 0.00778094632551074, + 0.004576831590384245, + 0.004231822211295366, + 0.002859804779291153, + 0.004866960924118757, + 0.004172820597887039, + 0.0019564847461879253, + 0.0022918570321053267, + 0.003028718987479806, + 0.0025932525750249624, + 0.008351054973900318, + 0.008960387669503689, + 0.008288788609206676, + 0.00733506353572011, + 0.004547470714896917, + 0.004199881572276354, + 0.006266520358622074, + 0.005113649647682905, + 0.008617773652076721, + 0.010257882066071033, + 0.011323448270559311, + 0.016838163137435913, + 0.0038002824876457453, + 0.004264386370778084, + 0.0059087700210511684, + 0.005633723922073841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.09981244802474976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902248948812485 + ], + "opposite_logprob_tokens": [ + -2.5790224075317383 + ], + "logprob_reward": -0.07902248948812485, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902248948812485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36489367485046387, + 0.18081232905387878, + 0.04765073582530022, + 0.057645879685878754, + 0.043359458446502686, + 0.0422670878469944, + 0.03858666494488716, + 0.12685571610927582, + 0.06163257360458374, + 0.0362958163022995 + ], + "credit_map": [ + 0.03385326638817787, + 0.03705888241529465, + 0.06911075860261917, + 0.10526861995458603, + 0.02883978560566902, + 0.028129611164331436, + 0.03553527221083641, + 0.02709750458598137, + 0.014067821204662323, + 0.023997245356440544, + 0.027131037786602974, + 0.026750031858682632, + 0.021096469834446907, + 0.017394592985510826, + 0.027120940387248993, + 0.023254213854670525, + 0.004241021350026131, + 0.006794792599976063, + 0.006498485803604126, + 0.006534623447805643, + 0.0059290602803230286, + 0.0052584270015358925, + 0.006900038570165634, + 0.00549429003149271, + 0.006404327694326639, + 0.006324117071926594, + 0.007711285725235939, + 0.009369575418531895, + 0.007642663549631834, + 0.0080767422914505, + 0.006142446771264076, + 0.005974726285785437, + 0.00512530654668808, + 0.003889968618750572, + 0.005932806525379419, + 0.0077535174787044525, + 0.005652354098856449, + 0.004696072079241276, + 0.004412675276398659, + 0.005896763410419226, + 0.005113547667860985, + 0.005943817552179098, + 0.006992271635681391, + 0.01069346722215414, + 0.003130903234705329, + 0.002728578634560108, + 0.004358113277703524, + 0.0033063937444239855, + 0.005686956923455, + 0.004206255078315735, + 0.005292351357638836, + 0.009182632900774479, + 0.0031322280410677195, + 0.0030491624493151903, + 0.004666964057832956, + 0.003370118560269475, + 0.02731022983789444, + 0.007708342280238867, + 0.04346040263772011, + 0.017523162066936493, + 0.0050503103993833065, + 0.006668988149613142, + 0.007380046881735325, + 0.011754232458770275, + 0.007183467969298363, + 0.011287492699921131, + 0.007907180115580559, + 0.01048206351697445, + 0.004489530809223652, + 0.0047221314162015915, + 0.009060130454599857, + 0.00650058314204216, + 0.004113644361495972, + 0.005606881342828274, + 0.0060071866028010845, + 0.009416991844773293, + 0.002444320125505328, + 0.0021976542193442583, + 0.0035372735001146793, + 0.0029718675650656223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.07902248948812485 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643122389912605 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643122389912605, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643122389912605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2276497483253479, + 0.11455009132623672, + 0.04593569040298462, + 0.04772000014781952, + 0.15111351013183594, + 0.049061667174100876, + 0.06365237385034561, + 0.03514406830072403, + 0.22192752361297607, + 0.043245330452919006 + ], + "credit_map": [ + 0.024343101307749748, + 0.03416920080780983, + 0.03513450175523758, + 0.03569481521844864, + 0.03431329503655434, + 0.018619205802679062, + 0.025382589548826218, + 0.01999303698539734, + 0.010014926083385944, + 0.01160335075110197, + 0.016933824867010117, + 0.022790344431996346, + 0.0088580921292305, + 0.011113550513982773, + 0.010532601736485958, + 0.022703399881720543, + 0.005504767410457134, + 0.006454091519117355, + 0.00755808362737298, + 0.010468131862580776, + 0.003535232041031122, + 0.00459688575938344, + 0.0034312941133975983, + 0.004387205000966787, + 0.004521510098129511, + 0.007480586878955364, + 0.007952684536576271, + 0.012943782843649387, + 0.00303387432359159, + 0.003209925489500165, + 0.00555287254974246, + 0.003024765057489276, + 0.018003512173891068, + 0.021688375622034073, + 0.021047500893473625, + 0.022730771452188492, + 0.017535530030727386, + 0.016530832275748253, + 0.016703611239790916, + 0.01687336526811123, + 0.005444783251732588, + 0.006135273724794388, + 0.008466443978250027, + 0.008216812275350094, + 0.0052254642359912395, + 0.004508309066295624, + 0.0060032145120203495, + 0.005061367992311716, + 0.008249056525528431, + 0.00725551787763834, + 0.010254676453769207, + 0.01532790157943964, + 0.004216952249407768, + 0.0040673003531992435, + 0.008337452076375484, + 0.005943513475358486, + 0.0044492692686617374, + 0.003328554565086961, + 0.008289925754070282, + 0.005794264376163483, + 0.002937789773568511, + 0.0032395701855421066, + 0.0038856903556734324, + 0.0032190040219575167, + 0.038711003959178925, + 0.021786438301205635, + 0.0337534174323082, + 0.024610182270407677, + 0.023653022944927216, + 0.021245867013931274, + 0.0357959158718586, + 0.0223716851323843, + 0.005090790800750256, + 0.00521700968965888, + 0.005819713696837425, + 0.0048542264848947525, + 0.0056836879812181, + 0.005457580089569092, + 0.005660411901772022, + 0.005461911670863628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.008643122389912605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233079195022583 + ], + "opposite_logprob_tokens": [ + -0.8983079195022583 + ], + "logprob_reward": -0.5233079195022583, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233079195022583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11581908911466599, + 0.09146317094564438, + 0.03799004852771759, + 0.049195677042007446, + 0.15808063745498657, + 0.0732506737112999, + 0.0725410208106041, + 0.13954363763332367, + 0.19342036545276642, + 0.06869567185640335 + ], + "credit_map": [ + 0.011652928777039051, + 0.012912309728562832, + 0.018874358385801315, + 0.026149950921535492, + 0.009604310616850853, + 0.009120849892497063, + 0.015389024280011654, + 0.01211535930633545, + 0.00785033404827118, + 0.012479756958782673, + 0.011728266254067421, + 0.013747216202318668, + 0.0075033968314528465, + 0.009062583558261395, + 0.011433198116719723, + 0.0176584180444479, + 0.00405574357137084, + 0.005817283410578966, + 0.0052873361855745316, + 0.005212689284235239, + 0.0035639507696032524, + 0.004538412671536207, + 0.004368033725768328, + 0.005146600306034088, + 0.005031286273151636, + 0.009725913405418396, + 0.007597432937473059, + 0.010849774815142155, + 0.00429384084418416, + 0.003467586124315858, + 0.004742832854390144, + 0.003487006528303027, + 0.024267496541142464, + 0.016348252072930336, + 0.025275234133005142, + 0.034726109355688095, + 0.013486799784004688, + 0.013404788449406624, + 0.015207785181701183, + 0.015364177525043488, + 0.0076302713714540005, + 0.009057451039552689, + 0.00970184151083231, + 0.011408920399844646, + 0.007095776032656431, + 0.008102906867861748, + 0.011868557892739773, + 0.008384952321648598, + 0.009974036365747452, + 0.008171632885932922, + 0.012277306988835335, + 0.021404659375548363, + 0.004093734547495842, + 0.0042708562687039375, + 0.006065357942134142, + 0.0062834350392222404, + 0.01946260780096054, + 0.020640596747398376, + 0.021714329719543457, + 0.02788672409951687, + 0.008680406957864761, + 0.012769514694809914, + 0.013664182275533676, + 0.01472528651356697, + 0.023265961557626724, + 0.02644641324877739, + 0.041770678013563156, + 0.022766685113310814, + 0.014273542910814285, + 0.019976628944277763, + 0.026356564834713936, + 0.018563905730843544, + 0.007994416169822216, + 0.012603694573044777, + 0.011138547211885452, + 0.017073560506105423, + 0.00461231917142868, + 0.0047336723655462265, + 0.005066066049039364, + 0.0054733953438699245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.5233079195022583 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369971364736557 + ], + "opposite_logprob_tokens": [ + -3.4086997509002686 + ], + "logprob_reward": -0.03369971364736557, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369971364736557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18931517004966736, + 0.06664099544286728, + 0.0609637051820755, + 0.07490348070859909, + 0.13790561258792877, + 0.20520733296871185, + 0.0667259618639946, + 0.06598836183547974, + 0.09899789839982986, + 0.033351413905620575 + ], + "credit_map": [ + 0.018518012017011642, + 0.017774859443306923, + 0.03324185311794281, + 0.04871675372123718, + 0.0151421669870615, + 0.01699019782245159, + 0.020090343430638313, + 0.018840989097952843, + 0.006097191479057074, + 0.00840618647634983, + 0.010842020623385906, + 0.011574188247323036, + 0.005906001199036837, + 0.006589657161384821, + 0.007814817130565643, + 0.009410935454070568, + 0.006140506826341152, + 0.008044391870498657, + 0.008439428173005581, + 0.009923843666911125, + 0.006441918667405844, + 0.006528964266180992, + 0.006608480121940374, + 0.00883617252111435, + 0.0063773528672754765, + 0.014894760213792324, + 0.011036653071641922, + 0.01554127037525177, + 0.008668913505971432, + 0.005511509720236063, + 0.007908779196441174, + 0.004964246414601803, + 0.021786807104945183, + 0.01826660893857479, + 0.023266350850462914, + 0.025792434811592102, + 0.013292072340846062, + 0.012057075276970863, + 0.012429420836269855, + 0.011014841496944427, + 0.022038480266928673, + 0.02795344963669777, + 0.029195472598075867, + 0.03657348081469536, + 0.017882315441966057, + 0.021384546533226967, + 0.03177083283662796, + 0.018408773466944695, + 0.00803167000412941, + 0.011724427342414856, + 0.009642989374697208, + 0.01596720889210701, + 0.004823560360819101, + 0.006046160589903593, + 0.004882900044322014, + 0.005607055965811014, + 0.007470420561730862, + 0.009359755553305149, + 0.011688648723065853, + 0.008086095564067364, + 0.006924725603312254, + 0.005866781808435917, + 0.008945641107857227, + 0.007646303623914719, + 0.014742427505552769, + 0.01398521475493908, + 0.01977558434009552, + 0.014225522987544537, + 0.009630725719034672, + 0.007753661833703518, + 0.011097866110503674, + 0.007786902133375406, + 0.004379481542855501, + 0.006559425033628941, + 0.005310495384037495, + 0.005130232311785221, + 0.0026773102581501007, + 0.0035049058496952057, + 0.002701099030673504, + 0.0030884677544236183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.03369971364736557 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up transports the viewer into a surreal digital rendering of a dense, dwarf bamboo forest, set against a dark, invisible canopy. The bamboo's unusually broad, waxy leaves are a deep green, their surface gleaming with a damp sheen in the light, while their slender, segmented stems are sturdy. The entire scene is dramatically illuminated by a powerful beam of sunlight, cutting diagonally through the darkness. This light source is not constant; it flickers violently, casting the environment in rapid, stroboscopic pulses, alternating between dazzling light and deep shadows. With each golden flash, the bamboo is captured in a subtle, time-lapse-like movement, its leaves and stems delicately twisting and bending, intermittently toward the beam. In the shimmering air, tiny dust particles dance erratically, appearing and disappearing in the stroboscopic shafts of light, while shimmering dewdrops on the leaves catch each pulse, creating momentary, brilliant crystal flickers that shimmer with prismatic intensity, generating a captivating and intense atmosphere of energy.", + "seed": 1320, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000107/group_9/index_1096_seed_1320_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11793507842812687, + "vlm_logprob": -0.11793507842812687, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.11793507842812687, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07662513852119446 + ], + "opposite_logprob_tokens": [ + -2.701625108718872 + ], + "logprob_reward": -0.07662513852119446, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07662513852119446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17206138372421265, + 0.12134063988924026, + 0.1341894567012787, + 0.060294896364212036, + 0.07217156141996384, + 0.10105088353157043, + 0.15936075150966644, + 0.0665852501988411, + 0.051616769284009933, + 0.06132841482758522 + ], + "credit_map": [ + 0.014931386336684227, + 0.009782484732568264, + 0.05268125981092453, + 0.008380015380680561, + 0.008320778608322144, + 0.04268760606646538, + 0.015572141855955124, + 0.01970570720732212, + 0.01145036518573761, + 0.0061900862492620945, + 0.018579479306936264, + 0.008050018921494484, + 0.007528777699917555, + 0.028507737442851067, + 0.011380939744412899, + 0.029653232544660568, + 0.012550502084195614, + 0.006125192623585463, + 0.016522763296961784, + 0.010991234332323074, + 0.011329157277941704, + 0.02665656991302967, + 0.0178961381316185, + 0.03211790323257446, + 0.0075151012279093266, + 0.0038222665898501873, + 0.007066548801958561, + 0.0047776768915355206, + 0.00590794300660491, + 0.012408001348376274, + 0.00835457444190979, + 0.01044278685003519, + 0.01050633005797863, + 0.011149104684591293, + 0.007706822361797094, + 0.007351220585405827, + 0.008466390892863274, + 0.00974599551409483, + 0.00794711709022522, + 0.009298579767346382, + 0.008631236851215363, + 0.006930532865226269, + 0.011677497066557407, + 0.012285890989005566, + 0.013401081785559654, + 0.018948635086417198, + 0.012844194658100605, + 0.01633181795477867, + 0.011087683029472828, + 0.008197479881346226, + 0.01897302083671093, + 0.01231857668608427, + 0.007769734598696232, + 0.04126422852277756, + 0.01636420749127865, + 0.043385811150074005, + 0.010262065567076206, + 0.005054405424743891, + 0.010904037393629551, + 0.005577160511165857, + 0.008093085139989853, + 0.0075220875442028046, + 0.009460074827075005, + 0.009712335653603077, + 0.013554167002439499, + 0.003974472172558308, + 0.006441511679440737, + 0.0035844147205352783, + 0.003313918597996235, + 0.006383981090039015, + 0.00796264037489891, + 0.00640166224911809, + 0.01831173524260521, + 0.003613706910982728, + 0.010997605510056019, + 0.0029506823047995567, + 0.0038189049810171127, + 0.009516499936580658, + 0.005066084209829569, + 0.007053192704916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.07662513852119446 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011091836728155613 + ], + "opposite_logprob_tokens": [ + -4.511091709136963 + ], + "logprob_reward": -0.011091836728155613, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011091836728155613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13726550340652466, + 0.07000428438186646, + 0.09525909274816513, + 0.09870688617229462, + 0.08362945914268494, + 0.10854530334472656, + 0.04994076490402222, + 0.1690082997083664, + 0.11311013251543045, + 0.07453032582998276 + ], + "credit_map": [ + 0.022771773859858513, + 0.009725792333483696, + 0.030697589740157127, + 0.007208055350929499, + 0.008190270513296127, + 0.030180543661117554, + 0.0060071214102208614, + 0.022484343498945236, + 0.00744634959846735, + 0.003985196817666292, + 0.008992373012006283, + 0.008474117144942284, + 0.005422620102763176, + 0.01367811020463705, + 0.009249263443052769, + 0.012756248004734516, + 0.011338833719491959, + 0.0059916614554822445, + 0.010913790203630924, + 0.008401961997151375, + 0.011680320836603642, + 0.015582827851176262, + 0.01771838590502739, + 0.0136313047260046, + 0.013274343684315681, + 0.012354315258562565, + 0.009532200172543526, + 0.014664917252957821, + 0.011078636161983013, + 0.014418602921068668, + 0.01035300362855196, + 0.013030854053795338, + 0.00963396392762661, + 0.010468768887221813, + 0.00833041500300169, + 0.009606856852769852, + 0.012936572544276714, + 0.013096936978399754, + 0.009103260934352875, + 0.01045267190784216, + 0.008684678934514523, + 0.007022578734904528, + 0.0119996452704072, + 0.01105371955782175, + 0.01360448356717825, + 0.022350097075104713, + 0.017661118879914284, + 0.016168976202607155, + 0.004075153265148401, + 0.004462084732949734, + 0.00695678498595953, + 0.00406789593398571, + 0.005024418700486422, + 0.010376481339335442, + 0.005267585627734661, + 0.009710355661809444, + 0.02684108167886734, + 0.013125942088663578, + 0.0325535386800766, + 0.010203872807323933, + 0.014194302260875702, + 0.03949932008981705, + 0.013037662953138351, + 0.0195525661110878, + 0.046560321003198624, + 0.00769276637583971, + 0.01978464238345623, + 0.0062814150005578995, + 0.004821696784347296, + 0.011867712251842022, + 0.005225151777267456, + 0.010876413434743881, + 0.027037769556045532, + 0.004367119167000055, + 0.013706103898584843, + 0.0028979224152863026, + 0.003509881440550089, + 0.011487451381981373, + 0.003450416261330247, + 0.008073654025793076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.011091836728155613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013410394312813878 + ], + "opposite_logprob_tokens": [ + -6.626340866088867 + ], + "logprob_reward": -0.0013410394312813878, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013410394312813878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27320143580436707, + 0.034862492233514786, + 0.353039413690567, + 0.04087723419070244, + 0.07257667183876038, + 0.052295755594968796, + 0.035342153161764145, + 0.07353206723928452, + 0.041485659778118134, + 0.022787103429436684 + ], + "credit_map": [ + 0.03234994038939476, + 0.013775521889328957, + 0.08865807205438614, + 0.0148356594145298, + 0.011723525822162628, + 0.056471243500709534, + 0.01351182535290718, + 0.04187566041946411, + 0.005471186712384224, + 0.002135484479367733, + 0.003967787604779005, + 0.004541672300547361, + 0.0025192738976329565, + 0.005345792509615421, + 0.004281980451196432, + 0.00659931218251586, + 0.019779888913035393, + 0.014108738861978054, + 0.01962881349027157, + 0.014071307145059109, + 0.014566229656338692, + 0.1303916573524475, + 0.022620845586061478, + 0.11787192523479462, + 0.006240540649741888, + 0.005748209543526173, + 0.00504737813025713, + 0.006198005750775337, + 0.0036360740195959806, + 0.005774911027401686, + 0.0038620366249233484, + 0.004370078444480896, + 0.009853646159172058, + 0.011263872496783733, + 0.009314912371337414, + 0.0067586335353553295, + 0.010889306664466858, + 0.00949724018573761, + 0.00761563191190362, + 0.007383420597761869, + 0.004640794824808836, + 0.0038538887165486813, + 0.006331492215394974, + 0.006039767060428858, + 0.007954956963658333, + 0.007428063079714775, + 0.008204305544495583, + 0.007842487655580044, + 0.003547435160726309, + 0.0037545382510870695, + 0.005187536124140024, + 0.00389465456828475, + 0.0035773657727986574, + 0.005856349598616362, + 0.004529921337962151, + 0.004994354676455259, + 0.011795585043728352, + 0.00545823760330677, + 0.013602815568447113, + 0.0030720150098204613, + 0.00507458858191967, + 0.019283076748251915, + 0.005509084556251764, + 0.009736658073961735, + 0.017394626513123512, + 0.003084055380895734, + 0.007338343188166618, + 0.0019042791100218892, + 0.0014852022286504507, + 0.004192101303488016, + 0.0019218376837670803, + 0.00416521355509758, + 0.005977396387606859, + 0.0016491326969116926, + 0.00412306934595108, + 0.0014257729053497314, + 0.0018370189936831594, + 0.0035879442002624273, + 0.0016553567256778479, + 0.0025314113590866327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.0013410394312813878 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.474271297454834 + ], + "opposite_logprob_tokens": [ + -0.974271297454834 + ], + "logprob_reward": -0.474271297454834, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.474271297454834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20752087235450745, + 0.052477143704891205, + 0.09484446793794632, + 0.0399259477853775, + 0.26073670387268066, + 0.1781402826309204, + 0.04394476115703583, + 0.03841991350054741, + 0.05359204113483429, + 0.030397824943065643 + ], + "credit_map": [ + 0.020354123786091805, + 0.010536962188780308, + 0.0679020956158638, + 0.010792460292577744, + 0.009619216434657574, + 0.05547884851694107, + 0.014465185813605785, + 0.018371988087892532, + 0.007871903479099274, + 0.0035708819050341845, + 0.006152219604700804, + 0.005792056210339069, + 0.004249090328812599, + 0.007975616492331028, + 0.006172855384647846, + 0.01069252472370863, + 0.010585103183984756, + 0.004671168513596058, + 0.009367629885673523, + 0.006410976406186819, + 0.008324914611876011, + 0.023795928806066513, + 0.01313003059476614, + 0.018558720126748085, + 0.005211139563471079, + 0.004045653622597456, + 0.0039534498937428, + 0.003808607580140233, + 0.0034148513805121183, + 0.009289604611694813, + 0.0036306711845099926, + 0.006571968551725149, + 0.0418379008769989, + 0.052176930010318756, + 0.02916826121509075, + 0.025737013667821884, + 0.03000912442803383, + 0.03032473474740982, + 0.02405289188027382, + 0.02742987871170044, + 0.015238740481436253, + 0.011092047207057476, + 0.023669537156820297, + 0.023009665310382843, + 0.021848006173968315, + 0.028637349605560303, + 0.023595567792654037, + 0.03104938566684723, + 0.003975987434387207, + 0.0025894504506140947, + 0.0048023550771176815, + 0.0028490095864981413, + 0.0021212364081293344, + 0.011977889575064182, + 0.0038957269862294197, + 0.011733109131455421, + 0.0045584989711642265, + 0.0022563443053513765, + 0.007662482094019651, + 0.0022775535471737385, + 0.003870920743793249, + 0.009873765520751476, + 0.003951971419155598, + 0.003968380391597748, + 0.018980640918016434, + 0.0042170253582298756, + 0.008794636465609074, + 0.002584639471024275, + 0.0027374497149139643, + 0.007027853280305862, + 0.0028463532216846943, + 0.006403444334864616, + 0.010308102704584599, + 0.0026293352711945772, + 0.005061605013906956, + 0.0017380430363118649, + 0.001780480146408081, + 0.0037170930299907923, + 0.0021628523245453835, + 0.0030003152787685394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.474271297454834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a dense forest of dwarf bamboo with unusually broad, waxy leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bamboo stems slender, segmented, and sturdy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by a powerful beam of sunlight cutting diagonally through a dark canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the light source flicker violently, creating rapid stroboscopic pulses between dazzling light and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the bamboo leaves and stems twist and bend intermittently toward the light beam in a time-lapse-like motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346080005168915 + ], + "opposite_logprob_tokens": [ + -3.65134596824646 + ], + "logprob_reward": -0.026346080005168915, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346080005168915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2704436480998993, + 0.07697880268096924, + 0.08899345993995667, + 0.06045536696910858, + 0.09800905734300613, + 0.14925944805145264, + 0.06086486205458641, + 0.088136687874794, + 0.07045838981866837, + 0.03640026971697807 + ], + "credit_map": [ + 0.029359279200434685, + 0.01855890080332756, + 0.077092245221138, + 0.016340864822268486, + 0.014815800823271275, + 0.05194355919957161, + 0.026051117107272148, + 0.03628189489245415, + 0.011921858415007591, + 0.004060844425112009, + 0.009960569441318512, + 0.005661756731569767, + 0.0068580652587115765, + 0.013598345220088959, + 0.012275437824428082, + 0.012641922570765018, + 0.008454789407551289, + 0.007214288227260113, + 0.009034914895892143, + 0.00628831097856164, + 0.011546670459210873, + 0.016586976125836372, + 0.01221621222794056, + 0.01765129715204239, + 0.008908330462872982, + 0.009322094731032848, + 0.0069429874420166016, + 0.010383382439613342, + 0.006284025963395834, + 0.0065099201165139675, + 0.006308170035481453, + 0.005796455312520266, + 0.01620129868388176, + 0.009765993803739548, + 0.01195011381059885, + 0.01008598692715168, + 0.01257765106856823, + 0.012422657571732998, + 0.01115963514894247, + 0.013845727778971195, + 0.014270723797380924, + 0.010117657482624054, + 0.01974300853908062, + 0.017935875803232193, + 0.011744479648768902, + 0.03167556971311569, + 0.015788627788424492, + 0.02798350527882576, + 0.006420142482966185, + 0.004179276991635561, + 0.00805885810405016, + 0.00529198395088315, + 0.005093297455459833, + 0.012805366888642311, + 0.007693291176110506, + 0.011322642676532269, + 0.011780517175793648, + 0.008366335183382034, + 0.016142375767230988, + 0.00621359096840024, + 0.011143466457724571, + 0.011250591836869717, + 0.014366964809596539, + 0.008872851729393005, + 0.027456140145659447, + 0.006461856886744499, + 0.012131725437939167, + 0.0034929190296679735, + 0.0030210658442229033, + 0.006397502031177282, + 0.003540897509083152, + 0.007956287823617458, + 0.012677418999373913, + 0.00257312785834074, + 0.00610303645953536, + 0.0021678023040294647, + 0.001806930755265057, + 0.004540537018328905, + 0.0029238576535135508, + 0.0036075578536838293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.026346080005168915 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_0/index_tensor([4261, 4261])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_0/index_tensor([4261, 4261])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5b8fbf5a0beac782f3e8533e586303a00368bc17 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_0/index_tensor([4261, 4261])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a slightly elevated angle, slowly pans to the right as a bus, its paint peeling and slightly worn, lurches to a halt on the curb. Finally, it stabilizes as the pneumatic doors hiss open. The first thing that comes into view is a hand gripping the door's metal armrest, polished to a shine by countless hands. Then, a weary woman in faded blue canvas overalls appears in the doorway. Her black hair is casually tied into a loose ponytail, a few strands clinging to her forehead with sweat. Her face is marked by a deep weariness, her eyelids heavy and her gaze dim. The dim light inside the bus highlights her slightly slumped shoulders, contrasting with the cool dusk outside. Carefully, she places one heavy, dusty work boot on the ground, then, with a heavy, almost dragging hand, descends the last step. The asphalt beneath her feet is damp from the recent rain, reflecting the scattered streetlights in the distance. As her weight completely left the bus, the bus body swayed slightly. The whole picture was filled with a quiet yet deep atmosphere of fatigue after get off work, with the interweaving of cold blue and warm yellow light and shadow.", + "seed": 6135, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_0/index_4261_seed_6135_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03333612447604537, + "vlm_logprob": -0.03333612447604537, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.03333612447604537, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1071343943476677 + ], + "opposite_logprob_tokens": [ + -2.7321343421936035 + ], + "logprob_reward": -0.1071343943476677, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.1071343943476677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5895770192146301, + 0.14017841219902039, + 0.05102202296257019, + 0.021836120635271072, + 0.021784864366054535, + 0.032540690153837204, + 0.030220482498407364, + 0.03151056170463562, + 0.0509391725063324, + 0.030390674248337746 + ], + "credit_map": [ + 0.11227767169475555, + 0.039290741086006165, + 0.10842234641313553, + 0.2270319163799286, + 0.01755930669605732, + 0.01825684681534767, + 0.04665577411651611, + 0.020082397386431694, + 0.012344300746917725, + 0.004125040490180254, + 0.004087250214070082, + 0.0041366987861692905, + 0.0452478751540184, + 0.00655975379049778, + 0.054925285279750824, + 0.008752209134399891, + 0.005024616606533527, + 0.0081355981528759, + 0.005691696424037218, + 0.005448959302157164, + 0.004437900148332119, + 0.006533341482281685, + 0.008544898591935635, + 0.007205009460449219, + 0.002149130217730999, + 0.0049664173275232315, + 0.002204708755016327, + 0.003676320891827345, + 0.0022816378623247147, + 0.0021750566083937883, + 0.0025439986493438482, + 0.0018388486932963133, + 0.004131566733121872, + 0.001963748363777995, + 0.0027691710274666548, + 0.0030389055609703064, + 0.0027517585549503565, + 0.0015627334360033274, + 0.002171875210478902, + 0.003395107574760914, + 0.002411909867078066, + 0.005661167204380035, + 0.0033800839446485043, + 0.004029361996799707, + 0.006453494541347027, + 0.0033877422101795673, + 0.0036453178618103266, + 0.0035716129932552576, + 0.005555558949708939, + 0.003960130736231804, + 0.003813541727140546, + 0.00417364202439785, + 0.0038122818805277348, + 0.0019935695454478264, + 0.00436119269579649, + 0.002550567267462611, + 0.0041114091873168945, + 0.003677910193800926, + 0.003704883623868227, + 0.004273887723684311, + 0.004086278844624758, + 0.004145227372646332, + 0.0044944267719984055, + 0.0030165393836796284, + 0.006121812388300896, + 0.004822385031729937, + 0.005437821615487337, + 0.009550492279231548, + 0.006409005261957645, + 0.005067997612059116, + 0.009783733636140823, + 0.0037459253799170256, + 0.007828382775187492, + 0.003482847474515438, + 0.006735473871231079, + 0.0026355977170169353, + 0.002069594105705619, + 0.0015120403841137886, + 0.0032979464158415794, + 0.002828790806233883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.1071343943476677 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012565725483000278 + ], + "opposite_logprob_tokens": [ + -4.387565612792969 + ], + "logprob_reward": -0.012565725483000278, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012565725483000278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.508581817150116, + 0.06212402507662773, + 0.11671212315559387, + 0.033606596291065216, + 0.045134786516427994, + 0.06900167465209961, + 0.039007045328617096, + 0.015807272866368294, + 0.07966217398643494, + 0.030362490564584732 + ], + "credit_map": [ + 0.07283217459917068, + 0.03135339915752411, + 0.10961771756410599, + 0.20402124524116516, + 0.016721246764063835, + 0.014818597584962845, + 0.040117792785167694, + 0.019099658355116844, + 0.0063825687393546104, + 0.0030093311797827482, + 0.00374723132699728, + 0.0030120054725557566, + 0.019670702517032623, + 0.003997009713202715, + 0.016983527690172195, + 0.005321649368852377, + 0.011224808171391487, + 0.022583575919270515, + 0.011057171039283276, + 0.013157560490071774, + 0.012636896222829819, + 0.0133878318592906, + 0.017592353746294975, + 0.015071926638484001, + 0.0033601978793740273, + 0.003891589120030403, + 0.002503906609490514, + 0.005223860498517752, + 0.005195749923586845, + 0.004660567734390497, + 0.004037092439830303, + 0.004733633249998093, + 0.006578576285392046, + 0.004109599161893129, + 0.005181671120226383, + 0.006611976772546768, + 0.005322656594216824, + 0.004255572333931923, + 0.005399104207754135, + 0.007675632834434509, + 0.009862481616437435, + 0.006727934814989567, + 0.007753878831863403, + 0.014152450487017632, + 0.008203836157917976, + 0.00399972265586257, + 0.012690501287579536, + 0.00561086880043149, + 0.0065889740362763405, + 0.00543650658801198, + 0.004635652992874384, + 0.007069769781082869, + 0.004029384348541498, + 0.0028995275497436523, + 0.005414699204266071, + 0.002932528732344508, + 0.0015801297267898917, + 0.0015207104152068496, + 0.0019841883331537247, + 0.002441035583615303, + 0.0019876686856150627, + 0.00245524849742651, + 0.0024454959202557802, + 0.0013927944237366319, + 0.006559946574270725, + 0.0049044047482311726, + 0.023979824036359787, + 0.023532193154096603, + 0.004754556808620691, + 0.004858169239014387, + 0.007924466393887997, + 0.0031486155930906534, + 0.007494042627513409, + 0.004009874537587166, + 0.005655363202095032, + 0.002795297186821699, + 0.0022219843231141567, + 0.0019623900298029184, + 0.0033820290118455887, + 0.0028415110427886248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.012565725483000278 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643949404358864 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008643949404358864, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643949404358864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4864254295825958, + 0.07295780628919601, + 0.030718635767698288, + 0.040855202823877335, + 0.033752575516700745, + 0.06280599534511566, + 0.12173250317573547, + 0.017145689576864243, + 0.04912789165973663, + 0.08447834849357605 + ], + "credit_map": [ + 0.0834038257598877, + 0.032563816756010056, + 0.09640246629714966, + 0.19051975011825562, + 0.015270048752427101, + 0.013529987074434757, + 0.03937655687332153, + 0.015358974225819111, + 0.006381096318364143, + 0.002727582585066557, + 0.0028854890260845423, + 0.0026160040870308876, + 0.02291058748960495, + 0.004352052696049213, + 0.026265211403369904, + 0.004819783847779036, + 0.004836008884012699, + 0.0042527783662080765, + 0.0031623304821550846, + 0.003291608765721321, + 0.002628283342346549, + 0.005480692721903324, + 0.0033312945161014795, + 0.0037356384564191103, + 0.004553013015538454, + 0.008130521513521671, + 0.003743537003174424, + 0.004660055972635746, + 0.007667925208806992, + 0.0032382667995989323, + 0.004978991113603115, + 0.0038828898686915636, + 0.005733526311814785, + 0.0029826064128428698, + 0.004048188216984272, + 0.005015812814235687, + 0.004113312810659409, + 0.002639901591464877, + 0.0035746076609939337, + 0.005644617602229118, + 0.0064475685358047485, + 0.0060419621877372265, + 0.0077602616511285305, + 0.014629454351961613, + 0.007516537327319384, + 0.004099476151168346, + 0.012641149573028088, + 0.0036695818416774273, + 0.024193175137043, + 0.02557477541267872, + 0.016133196651935577, + 0.012453053146600723, + 0.010733689181506634, + 0.010917182080447674, + 0.010430198162794113, + 0.011297235265374184, + 0.0017063138075172901, + 0.0019721644930541515, + 0.001824775943532586, + 0.0020609640050679445, + 0.0022532125003635883, + 0.0025639499071985483, + 0.0021524152252823114, + 0.0026118922978639603, + 0.005558178760111332, + 0.005331335123628378, + 0.005561192985624075, + 0.00788226630538702, + 0.006372442934662104, + 0.007669248152524233, + 0.007627897895872593, + 0.0031253264751285315, + 0.017420027405023575, + 0.014004983939230442, + 0.014489095658063889, + 0.006844600196927786, + 0.007019719108939171, + 0.006132789421826601, + 0.011651145294308662, + 0.006915986072272062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.008643949404358864 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033699944615364075 + ], + "opposite_logprob_tokens": [ + -3.4086999893188477 + ], + "logprob_reward": -0.033699944615364075, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033699944615364075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3700445592403412, + 0.09123384952545166, + 0.13900768756866455, + 0.03566331788897514, + 0.02433398738503456, + 0.06588179618120193, + 0.09342700988054276, + 0.02735901065170765, + 0.11253898590803146, + 0.04050980880856514 + ], + "credit_map": [ + 0.07250852137804031, + 0.02147422730922699, + 0.05722422897815704, + 0.12232468277215958, + 0.019543606787919998, + 0.02640322782099247, + 0.032807979732751846, + 0.01775810495018959, + 0.007165293209254742, + 0.004456227645277977, + 0.003974351100623608, + 0.004151429980993271, + 0.02822370082139969, + 0.005635254085063934, + 0.030060531571507454, + 0.007567060180008411, + 0.018026867881417274, + 0.025944704189896584, + 0.016804277896881104, + 0.01474856212735176, + 0.013109525665640831, + 0.015606307424604893, + 0.01705259643495083, + 0.017714839428663254, + 0.004145364742726088, + 0.005064534954726696, + 0.003252404974773526, + 0.004810012876987457, + 0.005869068671017885, + 0.004066197667270899, + 0.0038836519233882427, + 0.004572083242237568, + 0.003884982317686081, + 0.0024563376791775227, + 0.003972797188907862, + 0.003163754940032959, + 0.002779226051643491, + 0.0023673686664551497, + 0.0029792292043566704, + 0.0027302922680974007, + 0.003924360498785973, + 0.007067405618727207, + 0.005348231177777052, + 0.01584002375602722, + 0.009626652114093304, + 0.005544198211282492, + 0.013708455488085747, + 0.004822470713406801, + 0.012428045272827148, + 0.017956068739295006, + 0.01084208395332098, + 0.015214159153401852, + 0.01044167298823595, + 0.00779351219534874, + 0.00961807370185852, + 0.009133393876254559, + 0.0031472041737288237, + 0.003939615096896887, + 0.0033017999958246946, + 0.002904437016695738, + 0.00299944169819355, + 0.0032355296425521374, + 0.00398031622171402, + 0.0038506665732711554, + 0.015039779245853424, + 0.005504363216459751, + 0.03822913393378258, + 0.019867725670337677, + 0.0068774097599089146, + 0.016198959201574326, + 0.007268005982041359, + 0.00355360540561378, + 0.010135316289961338, + 0.004897417966276407, + 0.00767214410007, + 0.003549612360075116, + 0.002760361647233367, + 0.00301507581025362, + 0.004345329478383064, + 0.004134548828005791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.033699944615364075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636608529835939 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636608529835939, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636608529835939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47571849822998047, + 0.1018373966217041, + 0.05629430338740349, + 0.0351850651204586, + 0.06075083836913109, + 0.06650910526514053, + 0.06117227301001549, + 0.025226404890418053, + 0.07752330601215363, + 0.03978281468153 + ], + "credit_map": [ + 0.08657911419868469, + 0.031206658110022545, + 0.08169080317020416, + 0.1757403314113617, + 0.01662057638168335, + 0.02384955994784832, + 0.03590671345591545, + 0.02412472851574421, + 0.009200163185596466, + 0.004548873286694288, + 0.003832658054307103, + 0.0037760278210043907, + 0.034549325704574585, + 0.005631625652313232, + 0.03387370705604553, + 0.006425006780773401, + 0.007358422037214041, + 0.008609166368842125, + 0.006945371627807617, + 0.007911868393421173, + 0.0036762189120054245, + 0.008256187662482262, + 0.006282874848693609, + 0.007254182361066341, + 0.004297312814742327, + 0.004486250691115856, + 0.005238711833953857, + 0.005432948470115662, + 0.003923471551388502, + 0.0034205380361527205, + 0.004162141587585211, + 0.004223688971251249, + 0.012011473067104816, + 0.00552747817710042, + 0.009703250601887703, + 0.010353326797485352, + 0.006105482112616301, + 0.003540697041898966, + 0.005485025234520435, + 0.008024098351597786, + 0.005628359969705343, + 0.008022681809961796, + 0.0058507416397333145, + 0.013860883191227913, + 0.009703591465950012, + 0.00585512351244688, + 0.011950871907174587, + 0.00563684431836009, + 0.012929393909871578, + 0.007099924609065056, + 0.00787306483834982, + 0.007489430718123913, + 0.007056280970573425, + 0.003550795605406165, + 0.009303516708314419, + 0.005869857966899872, + 0.0027817205991595984, + 0.0028978134505450726, + 0.0028712020721286535, + 0.00447088060900569, + 0.003684493713080883, + 0.0022953711450099945, + 0.004213582258671522, + 0.002011338947340846, + 0.006928012706339359, + 0.00873887911438942, + 0.01836884208023548, + 0.016470160335302353, + 0.00605115108191967, + 0.008198641240596771, + 0.008905000984668732, + 0.003862610086798668, + 0.009178640320897102, + 0.0038875744212418795, + 0.00895605143159628, + 0.0036221537739038467, + 0.0028607582207769156, + 0.002332834992557764, + 0.005183922126889229, + 0.003760877763852477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.004636608529835939 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly elevated angle, slowly pans to the right as a bus, its paint peeling and slightly worn, lurches to a halt on the curb. Finally, it stabilizes as the pneumatic doors hiss open. The first thing that comes into view is a hand gripping the door's metal armrest, polished to a shine by countless hands. Then, a weary woman in faded blue canvas overalls appears in the doorway. Her black hair is casually tied into a loose ponytail, a few strands clinging to her forehead with sweat. Her face is marked by a deep weariness, her eyelids heavy and her gaze dim. The dim light inside the bus highlights her slightly slumped shoulders, contrasting with the cool dusk outside. Carefully, she places one heavy, dusty work boot on the ground, then, with a heavy, almost dragging hand, descends the last step. The asphalt beneath her feet is damp from the recent rain, reflecting the scattered streetlights in the distance. As her weight completely left the bus, the bus body swayed slightly. The whole picture was filled with a quiet yet deep atmosphere of fatigue after get off work, with the interweaving of cold blue and warm yellow light and shadow.", + "seed": 6136, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_0/index_4261_seed_6136_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017558015557006, + "vlm_logprob": -0.017558015557006, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.017558015557006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03875817358493805 + ], + "opposite_logprob_tokens": [ + -3.9137582778930664 + ], + "logprob_reward": -0.03875817358493805, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03875817358493805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07645846158266068, + 0.15251389145851135, + 0.1728050261735916, + 0.17154207825660706, + 0.25789380073547363, + 0.05304870009422302, + 0.026222500950098038, + 0.053949639201164246, + 0.014627116732299328, + 0.020938815549016 + ], + "credit_map": [ + 0.007509945426136255, + 0.0075281839817762375, + 0.007585704326629639, + 0.01371181569993496, + 0.011191537603735924, + 0.007973698899149895, + 0.01236629020422697, + 0.008591285906732082, + 0.010825620964169502, + 0.031088272109627724, + 0.005059096962213516, + 0.00885545089840889, + 0.013384563848376274, + 0.05381233990192413, + 0.022945405915379524, + 0.0065431417897343636, + 0.010829674080014229, + 0.021311752498149872, + 0.008459438569843769, + 0.016539601609110832, + 0.04412905499339104, + 0.03098250739276409, + 0.02416260354220867, + 0.01639038324356079, + 0.012406029738485813, + 0.03054378367960453, + 0.01049298420548439, + 0.01568843238055706, + 0.0344562791287899, + 0.02894841693341732, + 0.02092953957617283, + 0.01807660609483719, + 0.02865723706781864, + 0.03822356089949608, + 0.02236136794090271, + 0.033407408744096756, + 0.032145243138074875, + 0.0386628694832325, + 0.02958580106496811, + 0.03485029563307762, + 0.005453211255371571, + 0.006379913538694382, + 0.009767218492925167, + 0.006516957655549049, + 0.005500789266079664, + 0.007519767619669437, + 0.0058105215430259705, + 0.006100318394601345, + 0.004057331010699272, + 0.002345921704545617, + 0.003658173605799675, + 0.0033246451057493687, + 0.0034260316751897335, + 0.0038239918649196625, + 0.003068899968639016, + 0.002517508342862129, + 0.011223584413528442, + 0.006581200752407312, + 0.00981054361909628, + 0.00463629188016057, + 0.004726346582174301, + 0.00646708719432354, + 0.005899283103644848, + 0.0046052997931838036, + 0.0017323826905339956, + 0.0013600027887150645, + 0.0012353697093203664, + 0.0016342528397217393, + 0.0018389495089650154, + 0.002550288336351514, + 0.001665089512243867, + 0.002610781928524375, + 0.0020609728526324034, + 0.0020394001621752977, + 0.0027032624930143356, + 0.00299551198258996, + 0.0029538420494645834, + 0.0018328996375203133, + 0.004142691381275654, + 0.0022102349903434515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.03875817358493805 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636663056910038 + ], + "opposite_logprob_tokens": [ + -3.651366710662842 + ], + "logprob_reward": -0.02636663056910038, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636663056910038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14048929512500763, + 0.10092116892337799, + 0.256412148475647, + 0.10096762329339981, + 0.1305592656135559, + 0.08816775679588318, + 0.05055180564522743, + 0.06639379262924194, + 0.04290127009153366, + 0.022635867819190025 + ], + "credit_map": [ + 0.015366173349320889, + 0.013051637448370457, + 0.013605901971459389, + 0.02515493892133236, + 0.02244918793439865, + 0.015796169638633728, + 0.019801760092377663, + 0.015263527631759644, + 0.007282651960849762, + 0.022345134988427162, + 0.0041270023211836815, + 0.006573064252734184, + 0.011365966871380806, + 0.029637543484568596, + 0.014192122034728527, + 0.005397682543843985, + 0.012936769984662533, + 0.02624577470123768, + 0.011452032253146172, + 0.016821546480059624, + 0.06307941675186157, + 0.056609999388456345, + 0.040978092700242996, + 0.02828853577375412, + 0.007989896461367607, + 0.01443612203001976, + 0.008433960378170013, + 0.011612406931817532, + 0.017283592373132706, + 0.01756642386317253, + 0.011023391038179398, + 0.012621831148862839, + 0.013255303725600243, + 0.02132759615778923, + 0.008693738840520382, + 0.01419292576611042, + 0.019099485129117966, + 0.016544513404369354, + 0.017875023186206818, + 0.01957067847251892, + 0.014288963750004768, + 0.018076622858643532, + 0.013945324346423149, + 0.011254871264100075, + 0.007868001237511635, + 0.00681147538125515, + 0.007369447033852339, + 0.008553048595786095, + 0.004977119620889425, + 0.004261223133653402, + 0.006328668911010027, + 0.004248995333909988, + 0.006990276277065277, + 0.008396490477025509, + 0.007268136367201805, + 0.008080890402197838, + 0.01044787559658289, + 0.007510265335440636, + 0.006938501726835966, + 0.0064125340431928635, + 0.010731052607297897, + 0.008766861632466316, + 0.008246303535997868, + 0.007340394426137209, + 0.00398199213668704, + 0.00328381871804595, + 0.00426807114854455, + 0.00419551320374012, + 0.006747951731085777, + 0.006536821369081736, + 0.009197098203003407, + 0.0046900007873773575, + 0.002384036546573043, + 0.0028942711651325226, + 0.002946210792288184, + 0.002438184106722474, + 0.0033801391255110502, + 0.002207747893407941, + 0.003734417725354433, + 0.0026508623268455267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.02636663056910038 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254771463572979 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.01254771463572979, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254771463572979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06516336649656296, + 0.08914171159267426, + 0.0851927250623703, + 0.09387686848640442, + 0.07245409488677979, + 0.04680527374148369, + 0.10190866142511368, + 0.10665863007307053, + 0.2942157983779907, + 0.04458289220929146 + ], + "credit_map": [ + 0.007013483438640833, + 0.0045935166999697685, + 0.010669544339179993, + 0.009692954830825329, + 0.01134475041180849, + 0.0060719880275428295, + 0.010453051887452602, + 0.005324078723788261, + 0.005765172187238932, + 0.015412230975925922, + 0.004103057086467743, + 0.00794126745313406, + 0.013537345454096794, + 0.022032778710126877, + 0.014980357140302658, + 0.005369504913687706, + 0.004996051546186209, + 0.008401638828217983, + 0.004331060219556093, + 0.01068179216235876, + 0.021718503907322884, + 0.01602262258529663, + 0.012581733986735344, + 0.006459326483309269, + 0.005830729380249977, + 0.013762758113443851, + 0.0064360881224274635, + 0.009617030620574951, + 0.018275367096066475, + 0.014163791202008724, + 0.013106844387948513, + 0.012684259563684464, + 0.008298778906464577, + 0.009583142586052418, + 0.0041090031154453754, + 0.00685990322381258, + 0.012849170714616776, + 0.009132162667810917, + 0.011139079928398132, + 0.010482854209840298, + 0.007611737120896578, + 0.010279888287186623, + 0.004971250891685486, + 0.0055122459307312965, + 0.003659639274701476, + 0.004348390735685825, + 0.003998751286417246, + 0.006423370447009802, + 0.008408525958657265, + 0.011688238009810448, + 0.01590508595108986, + 0.024967513978481293, + 0.008074117824435234, + 0.014271344989538193, + 0.010055307298898697, + 0.008538525551557541, + 0.014877253212034702, + 0.013149854727089405, + 0.010501952841877937, + 0.0170820914208889, + 0.013749737292528152, + 0.013660755939781666, + 0.012229429557919502, + 0.011407546699047089, + 0.032480038702487946, + 0.027816375717520714, + 0.03773582726716995, + 0.032827407121658325, + 0.045074015855789185, + 0.03748494014143944, + 0.03135640546679497, + 0.04944077134132385, + 0.004033572506159544, + 0.004436194431036711, + 0.005467744078487158, + 0.006190687883645296, + 0.00556672690436244, + 0.004302049987018108, + 0.008385282009840012, + 0.006200634408742189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.01254771463572979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007629659026861191 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007629659026861191, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007629659026861191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13905346393585205, + 0.06832068413496017, + 0.13438144326210022, + 0.19782143831253052, + 0.04503336548805237, + 0.11742139607667923, + 0.07013396173715591, + 0.13943298161029816, + 0.0545773059129715, + 0.03382400795817375 + ], + "credit_map": [ + 0.013761917129158974, + 0.009985762648284435, + 0.015020894818007946, + 0.014821117743849754, + 0.025325264781713486, + 0.01663881726562977, + 0.02565954253077507, + 0.017840152606368065, + 0.005844594445079565, + 0.01425428781658411, + 0.005175698082894087, + 0.008462002500891685, + 0.01095440424978733, + 0.008453622460365295, + 0.010802245698869228, + 0.00437383446842432, + 0.008957518264651299, + 0.014835688285529613, + 0.006038374733179808, + 0.008597472682595253, + 0.03262478858232498, + 0.030137550085783005, + 0.01996959187090397, + 0.013220452703535557, + 0.019335901364684105, + 0.027793243527412415, + 0.024613700807094574, + 0.044723786413669586, + 0.018376655876636505, + 0.01698802039027214, + 0.022091738879680634, + 0.02389838919043541, + 0.004787919111549854, + 0.005976468324661255, + 0.005253439769148827, + 0.005538634955883026, + 0.004411268979310989, + 0.00874300580471754, + 0.004079638049006462, + 0.006242990959435701, + 0.013462179340422153, + 0.020186299458146095, + 0.01949239894747734, + 0.0144019341096282, + 0.011279422789812088, + 0.01008738111704588, + 0.014903167262673378, + 0.013608617708086967, + 0.008355407044291496, + 0.006519162096083164, + 0.007785856258124113, + 0.007467176765203476, + 0.011533787474036217, + 0.010358531959354877, + 0.012511571869254112, + 0.005602471996098757, + 0.0367007851600647, + 0.014821324497461319, + 0.013648984022438526, + 0.01254873163998127, + 0.015856444835662842, + 0.017817335203289986, + 0.015260335057973862, + 0.012779038399457932, + 0.005413324572145939, + 0.005362141877412796, + 0.007133719976991415, + 0.006320028565824032, + 0.00885842926800251, + 0.006622957531362772, + 0.007295301184058189, + 0.007571400608867407, + 0.003723553381860256, + 0.004166170489042997, + 0.006109622307121754, + 0.004789985250681639, + 0.004045791923999786, + 0.0032331966795027256, + 0.004199507646262646, + 0.003556176321581006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.007629659026861191 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024878999684005976 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.0024878999684005976, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024878999684005976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1313273161649704, + 0.09134220331907272, + 0.22455301880836487, + 0.06935793906450272, + 0.06122169271111488, + 0.0833560973405838, + 0.1029471904039383, + 0.12793537974357605, + 0.054098065942525864, + 0.05386114493012428 + ], + "credit_map": [ + 0.016660209745168686, + 0.011717753484845161, + 0.014986971393227577, + 0.014431282877922058, + 0.022583266720175743, + 0.0168848168104887, + 0.018391842022538185, + 0.01567116007208824, + 0.006415327545255423, + 0.016284814104437828, + 0.003792423289269209, + 0.006889100652188063, + 0.010950583033263683, + 0.027999917045235634, + 0.014876926317811012, + 0.004133096430450678, + 0.016767827793955803, + 0.02459115907549858, + 0.00985694769769907, + 0.014709224924445152, + 0.05636727064847946, + 0.04822922125458717, + 0.0348532572388649, + 0.019178107380867004, + 0.007049078121781349, + 0.010332354344427586, + 0.007766311522573233, + 0.013597049750387669, + 0.007887768559157848, + 0.007925061509013176, + 0.0067960359156131744, + 0.008004278875887394, + 0.0071318787522614, + 0.013383373618125916, + 0.0058900644071400166, + 0.008105268701910973, + 0.005545132793486118, + 0.00939357653260231, + 0.004344312474131584, + 0.007428079843521118, + 0.010888651013374329, + 0.012493813410401344, + 0.010641250759363174, + 0.010826432146131992, + 0.008617868646979332, + 0.011277931742370129, + 0.008230265229940414, + 0.010379878804087639, + 0.010794139467179775, + 0.008087356574833393, + 0.012667713686823845, + 0.008727284148335457, + 0.01516217365860939, + 0.0163009874522686, + 0.01590476930141449, + 0.015302762389183044, + 0.022846169769763947, + 0.018762897700071335, + 0.014814077876508236, + 0.01142956130206585, + 0.018087081611156464, + 0.015872398391366005, + 0.015913985669612885, + 0.010209194384515285, + 0.007696948014199734, + 0.005827903747558594, + 0.0069959042593836784, + 0.006488427519798279, + 0.0073317536152899265, + 0.006194877903908491, + 0.008172507397830486, + 0.005389742087572813, + 0.006200586445629597, + 0.005703740753233433, + 0.008371571078896523, + 0.007706067990511656, + 0.0073307533748447895, + 0.005439643748104572, + 0.007185233291238546, + 0.005923551507294178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.0024878999684005976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly elevated angle, slowly pans to the right as a bus, its paint peeling and slightly worn, lurches to a halt on the curb. Finally, it stabilizes as the pneumatic doors hiss open. The first thing that comes into view is a hand gripping the door's metal armrest, polished to a shine by countless hands. Then, a weary woman in faded blue canvas overalls appears in the doorway. Her black hair is casually tied into a loose ponytail, a few strands clinging to her forehead with sweat. Her face is marked by a deep weariness, her eyelids heavy and her gaze dim. The dim light inside the bus highlights her slightly slumped shoulders, contrasting with the cool dusk outside. Carefully, she places one heavy, dusty work boot on the ground, then, with a heavy, almost dragging hand, descends the last step. The asphalt beneath her feet is damp from the recent rain, reflecting the scattered streetlights in the distance. As her weight completely left the bus, the bus body swayed slightly. The whole picture was filled with a quiet yet deep atmosphere of fatigue after get off work, with the interweaving of cold blue and warm yellow light and shadow.", + "seed": 6137, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_0/index_4261_seed_6137_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019048252841457725, + "vlm_logprob": -0.019048252841457725, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.019048252841457725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06322778761386871 + ], + "opposite_logprob_tokens": [ + -3.563227891921997 + ], + "logprob_reward": -0.06322778761386871, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.06322778761386871, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23959097266197205, + 0.07192149758338928, + 0.05901680886745453, + 0.10949676483869553, + 0.1951666921377182, + 0.07351294159889221, + 0.0529567189514637, + 0.07357588410377502, + 0.07598888128995895, + 0.04877286031842232 + ], + "credit_map": [ + 0.029646310955286026, + 0.021648328751325607, + 0.05220546945929527, + 0.05849449709057808, + 0.016865020617842674, + 0.013280333951115608, + 0.030437376350164413, + 0.017013628035783768, + 0.013894475996494293, + 0.005496210418641567, + 0.014640463516116142, + 0.014763944782316685, + 0.004960430320352316, + 0.004671232309192419, + 0.006427044048905373, + 0.007067701313644648, + 0.004857067950069904, + 0.00665129441767931, + 0.008681230247020721, + 0.007806761655956507, + 0.0041991109028458595, + 0.012392515316605568, + 0.005114100873470306, + 0.00931472796946764, + 0.010094941593706608, + 0.011779512278735638, + 0.016636691987514496, + 0.012076225131750107, + 0.017943106591701508, + 0.009326329454779625, + 0.019952228292822838, + 0.011687727645039558, + 0.020311200991272926, + 0.026954542845487595, + 0.02330808900296688, + 0.03957011550664902, + 0.01972327008843422, + 0.020527200773358345, + 0.025582777336239815, + 0.01918948069214821, + 0.007406970951706171, + 0.005865368526428938, + 0.007705137599259615, + 0.010267142206430435, + 0.010225034318864346, + 0.010019275359809399, + 0.010354558937251568, + 0.011669455096125603, + 0.004451992455869913, + 0.006036240141838789, + 0.005263355560600758, + 0.006844053976237774, + 0.008413365110754967, + 0.0038945707492530346, + 0.013190691359341145, + 0.004862450994551182, + 0.010987335816025734, + 0.0052413782104849815, + 0.009130842983722687, + 0.0068350620567798615, + 0.012258498929440975, + 0.011124087497591972, + 0.009333176538348198, + 0.008665494620800018, + 0.006114635616540909, + 0.0066863237880170345, + 0.014465508051216602, + 0.010104391723871231, + 0.012417434714734554, + 0.007097954396158457, + 0.010099263861775398, + 0.009003371000289917, + 0.007542389910668135, + 0.004229268524795771, + 0.005413461942225695, + 0.006103888154029846, + 0.005420699715614319, + 0.007505071349442005, + 0.006049579009413719, + 0.0065085007809102535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.06322778761386871 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011095608584582806 + ], + "opposite_logprob_tokens": [ + -4.5110955238342285 + ], + "logprob_reward": -0.011095608584582806, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011095608584582806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1306084543466568, + 0.05737042799592018, + 0.07070622593164444, + 0.10735518485307693, + 0.1890212893486023, + 0.14793168008327484, + 0.13758814334869385, + 0.04740080237388611, + 0.06994388997554779, + 0.04207386076450348 + ], + "credit_map": [ + 0.014826205559074879, + 0.012128720059990883, + 0.02653476595878601, + 0.02273411862552166, + 0.01404822152107954, + 0.011711221188306808, + 0.01823878474533558, + 0.010386426001787186, + 0.00660886662080884, + 0.007269745226949453, + 0.01238901074975729, + 0.009505419060587883, + 0.005017941351979971, + 0.0051094419322907925, + 0.005180204287171364, + 0.006289796903729439, + 0.0066736117005348206, + 0.008773105219006538, + 0.011283617466688156, + 0.008940011262893677, + 0.007666247431188822, + 0.010051585733890533, + 0.00864717923104763, + 0.0086708664894104, + 0.007352120243012905, + 0.006171874701976776, + 0.01440098974853754, + 0.008782624267041683, + 0.01874271221458912, + 0.014950122684240341, + 0.020817989483475685, + 0.01613675430417061, + 0.017079723998904228, + 0.034943126142024994, + 0.018198831006884575, + 0.04884342476725578, + 0.01839364878833294, + 0.021768081933259964, + 0.01746349036693573, + 0.012330989353358746, + 0.01579202152788639, + 0.020320823416113853, + 0.021014723926782608, + 0.02182667702436447, + 0.019432764500379562, + 0.016277022659778595, + 0.017907997593283653, + 0.015359656885266304, + 0.011187613010406494, + 0.016115551814436913, + 0.018259849399328232, + 0.013204377144575119, + 0.01885746791958809, + 0.013219243846833706, + 0.024543650448322296, + 0.022200390696525574, + 0.006328430492430925, + 0.004208277445286512, + 0.0067675416357815266, + 0.005044633522629738, + 0.006108797155320644, + 0.007327761501073837, + 0.00529312901198864, + 0.0063222311437129974, + 0.00562139181420207, + 0.004167073406279087, + 0.00988957192748785, + 0.006654367782175541, + 0.011007413268089294, + 0.01114280242472887, + 0.009796044789254665, + 0.011665230616927147, + 0.003855025162920356, + 0.004201470408588648, + 0.005270853638648987, + 0.0039718556217849255, + 0.005341633222997189, + 0.008004480041563511, + 0.004675318021327257, + 0.006753227673470974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.011095608584582806 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786956012248993 + ], + "opposite_logprob_tokens": [ + -4.634787082672119 + ], + "logprob_reward": -0.009786956012248993, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786956012248993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09742788225412369, + 0.06897295266389847, + 0.05620391666889191, + 0.06302258372306824, + 0.09515988081693649, + 0.09292065352201462, + 0.15948402881622314, + 0.0691651701927185, + 0.17626623809337616, + 0.12137661874294281 + ], + "credit_map": [ + 0.015091556124389172, + 0.010840206407010555, + 0.013112693093717098, + 0.016867732629179955, + 0.008761588484048843, + 0.0104281110689044, + 0.012329229153692722, + 0.009996768087148666, + 0.011139296926558018, + 0.006528411991894245, + 0.016649501398205757, + 0.013120966963469982, + 0.004649930167943239, + 0.004497535992413759, + 0.006563288625329733, + 0.005824022926390171, + 0.004564284346997738, + 0.004624973516911268, + 0.009341047145426273, + 0.008526886813342571, + 0.006991873029619455, + 0.011402753181755543, + 0.006194845773279667, + 0.00455725472420454, + 0.006343484856188297, + 0.005983333103358746, + 0.009432115592062473, + 0.004595581442117691, + 0.009584291838109493, + 0.007119421381503344, + 0.010387825779616833, + 0.009576531127095222, + 0.0086683239787817, + 0.02126624993979931, + 0.011104735545814037, + 0.018810560926795006, + 0.00919287372380495, + 0.00923397857695818, + 0.010018073953688145, + 0.006865084636956453, + 0.009843508712947369, + 0.009484450332820415, + 0.010636841878294945, + 0.017902975901961327, + 0.011578052304685116, + 0.010924449190497398, + 0.011028526350855827, + 0.011521843262016773, + 0.013930069282650948, + 0.023753084242343903, + 0.015837647020816803, + 0.021683456376194954, + 0.03025355562567711, + 0.010065275244414806, + 0.03373560309410095, + 0.010225345380604267, + 0.00748252309858799, + 0.007546743378043175, + 0.00916292518377304, + 0.0115126371383667, + 0.011256212368607521, + 0.0071779172867536545, + 0.010844227857887745, + 0.004181980155408382, + 0.010768427513539791, + 0.019839493557810783, + 0.01730736345052719, + 0.03477167338132858, + 0.028676148504018784, + 0.017235778272151947, + 0.03042629547417164, + 0.01724105328321457, + 0.011945502832531929, + 0.014284534379839897, + 0.014314945787191391, + 0.011761309579014778, + 0.01765367202460766, + 0.012378709390759468, + 0.015047475695610046, + 0.023990463465452194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.009786956012248993 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641703985631466 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.008641703985631466, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641703985631466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2305782586336136, + 0.04370781406760216, + 0.08224563300609589, + 0.09774642437696457, + 0.06295944005250931, + 0.09839051961898804, + 0.07466412335634232, + 0.04442882910370827, + 0.2273784577846527, + 0.03790048137307167 + ], + "credit_map": [ + 0.0377686433494091, + 0.021008308976888657, + 0.05854581296443939, + 0.03999735414981842, + 0.015137030743062496, + 0.009534980170428753, + 0.035121891647577286, + 0.013464231044054031, + 0.007403271272778511, + 0.004773177672177553, + 0.007799865212291479, + 0.006324514746665955, + 0.0031201012898236513, + 0.004303720314055681, + 0.0038104418199509382, + 0.006172723136842251, + 0.00826267246156931, + 0.008399412967264652, + 0.011018357239663601, + 0.012416133657097816, + 0.008586538024246693, + 0.01784202642738819, + 0.008940153755247593, + 0.006780343130230904, + 0.006509855855256319, + 0.006824164185672998, + 0.012576653622090816, + 0.008770907297730446, + 0.016674550250172615, + 0.009056257084012032, + 0.019569888710975647, + 0.017764145508408546, + 0.005290981847792864, + 0.007476257160305977, + 0.0059748683124780655, + 0.009663174860179424, + 0.0055649373680353165, + 0.008915713988244534, + 0.01149086095392704, + 0.008582647889852524, + 0.01011833269149065, + 0.01451861672103405, + 0.010701117105782032, + 0.017273511737585068, + 0.012912111356854439, + 0.010568762198090553, + 0.011177696287631989, + 0.011120363138616085, + 0.0066140806302428246, + 0.010924629867076874, + 0.008574018254876137, + 0.008592848666012287, + 0.011517761275172234, + 0.006147470790892839, + 0.01100938767194748, + 0.011283932253718376, + 0.005435652565211058, + 0.0062737478874623775, + 0.0038366992957890034, + 0.004694903269410133, + 0.00696706073358655, + 0.00649099238216877, + 0.0053772046230733395, + 0.005352569278329611, + 0.017981087788939476, + 0.01716364361345768, + 0.031013906002044678, + 0.019902724772691727, + 0.034741222858428955, + 0.04193934053182602, + 0.026102956384420395, + 0.03853358328342438, + 0.005214059725403786, + 0.0025288539472967386, + 0.0041355714201927185, + 0.00430578226223588, + 0.005947384051978588, + 0.0058088405057787895, + 0.005338630173355341, + 0.004621358588337898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.008641703985631466 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002489208010956645 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.002489208010956645, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002489208010956645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2390449047088623, + 0.07688526064157486, + 0.10135150700807571, + 0.07700254023075104, + 0.07101187855005264, + 0.06824415922164917, + 0.13107357919216156, + 0.08139274269342422, + 0.08662001043558121, + 0.06737349927425385 + ], + "credit_map": [ + 0.04285714030265808, + 0.026665380224585533, + 0.028852589428424835, + 0.04946375638246536, + 0.020249735563993454, + 0.017711002379655838, + 0.036641232669353485, + 0.016604067757725716, + 0.015351977199316025, + 0.008386554196476936, + 0.01627240516245365, + 0.011350223794579506, + 0.005870149005204439, + 0.005536838434636593, + 0.006737278774380684, + 0.007379830814898014, + 0.009668145328760147, + 0.010993794538080692, + 0.012451672926545143, + 0.012359241023659706, + 0.011689319275319576, + 0.023392098024487495, + 0.01249237172305584, + 0.008304856717586517, + 0.006162465550005436, + 0.005650490056723356, + 0.012652439065277576, + 0.007910256274044514, + 0.014328999444842339, + 0.00876467302441597, + 0.013850823044776917, + 0.0076823849231004715, + 0.006863510236144066, + 0.013759372755885124, + 0.00622792961075902, + 0.015257258899509907, + 0.006414193194359541, + 0.00859606359153986, + 0.008463049307465553, + 0.005430495832115412, + 0.006686599459499121, + 0.009683942422270775, + 0.008048365823924541, + 0.011929971165955067, + 0.008198161609470844, + 0.007157184649258852, + 0.009113925509154797, + 0.007426001131534576, + 0.01181348878890276, + 0.01356747280806303, + 0.017962567508220673, + 0.013393399305641651, + 0.02092556469142437, + 0.011848430149257183, + 0.023336436599493027, + 0.018226204439997673, + 0.01024681143462658, + 0.010718410834670067, + 0.007325030397623777, + 0.0068643647246062756, + 0.011743209324777126, + 0.011403507553040981, + 0.01101195439696312, + 0.012079447507858276, + 0.007065669167786837, + 0.008779069408774376, + 0.01244644820690155, + 0.013286103494465351, + 0.013635371811687946, + 0.009160607121884823, + 0.013841334730386734, + 0.008405406959354877, + 0.010175477713346481, + 0.006127160042524338, + 0.007444326765835285, + 0.006361893843859434, + 0.006843981798738241, + 0.013946674764156342, + 0.00674933847039938, + 0.009724639356136322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.002489208010956645 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly elevated angle, slowly pans to the right as a bus, its paint peeling and slightly worn, lurches to a halt on the curb. Finally, it stabilizes as the pneumatic doors hiss open. The first thing that comes into view is a hand gripping the door's metal armrest, polished to a shine by countless hands. Then, a weary woman in faded blue canvas overalls appears in the doorway. Her black hair is casually tied into a loose ponytail, a few strands clinging to her forehead with sweat. Her face is marked by a deep weariness, her eyelids heavy and her gaze dim. The dim light inside the bus highlights her slightly slumped shoulders, contrasting with the cool dusk outside. Carefully, she places one heavy, dusty work boot on the ground, then, with a heavy, almost dragging hand, descends the last step. The asphalt beneath her feet is damp from the recent rain, reflecting the scattered streetlights in the distance. As her weight completely left the bus, the bus body swayed slightly. The whole picture was filled with a quiet yet deep atmosphere of fatigue after get off work, with the interweaving of cold blue and warm yellow light and shadow.", + "seed": 6138, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_0/index_4261_seed_6138_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021399027667939663, + "vlm_logprob": -0.021399027667939663, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.021399027667939663, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05988699197769165 + ], + "opposite_logprob_tokens": [ + -3.684886932373047 + ], + "logprob_reward": -0.05988699197769165, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.05988699197769165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15110374987125397, + 0.09231758862733841, + 0.07806267589330673, + 0.2738623321056366, + 0.05852646753191948, + 0.04345129802823067, + 0.04389502480626106, + 0.11339873820543289, + 0.04889844357967377, + 0.09648371487855911 + ], + "credit_map": [ + 0.020235955715179443, + 0.019053872674703598, + 0.01704907976090908, + 0.016333431005477905, + 0.0233549065887928, + 0.025304323062300682, + 0.015250515192747116, + 0.014521668665111065, + 0.013809116557240486, + 0.010920836590230465, + 0.010841049253940582, + 0.01092070247977972, + 0.010577751323580742, + 0.011902004480361938, + 0.01039937511086464, + 0.01294674351811409, + 0.008093979209661484, + 0.006158981006592512, + 0.009708386845886707, + 0.012327756732702255, + 0.011428039520978928, + 0.013598410412669182, + 0.009276852943003178, + 0.007470267824828625, + 0.011045241728425026, + 0.009179414249956608, + 0.07963310927152634, + 0.010398877784609795, + 0.03629984334111214, + 0.10781241208314896, + 0.008178885094821453, + 0.011314553208649158, + 0.005291812587529421, + 0.009727993980050087, + 0.004197534639388323, + 0.008438440039753914, + 0.008150560781359673, + 0.009827387519180775, + 0.0056831142865121365, + 0.007209617644548416, + 0.003973144572228193, + 0.0046888128854334354, + 0.005940466653555632, + 0.005351301748305559, + 0.006349519826471806, + 0.005586546845734119, + 0.006112458184361458, + 0.005449043586850166, + 0.004202302545309067, + 0.0046562450006604195, + 0.004527771379798651, + 0.005540824495255947, + 0.00599899236112833, + 0.00800164882093668, + 0.005778267979621887, + 0.005188970360904932, + 0.010437916964292526, + 0.018527861684560776, + 0.024470971897244453, + 0.01257705595344305, + 0.012838093563914299, + 0.01164866890758276, + 0.01069245021790266, + 0.012205718085169792, + 0.004366068635135889, + 0.004056669771671295, + 0.0054778470657765865, + 0.008155564777553082, + 0.006938389502465725, + 0.004365100525319576, + 0.008306026458740234, + 0.0072327726520597935, + 0.011134080588817596, + 0.013679325580596924, + 0.005486035253852606, + 0.008962460793554783, + 0.013875165022909641, + 0.025169886648654938, + 0.009802565909922123, + 0.008374189957976341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.05988699197769165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012565725483000278 + ], + "opposite_logprob_tokens": [ + -4.387565612792969 + ], + "logprob_reward": -0.012565725483000278, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012565725483000278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14728567004203796, + 0.1023055836558342, + 0.166365846991539, + 0.14664554595947266, + 0.08232761919498444, + 0.04373908415436745, + 0.07748742401599884, + 0.12455424666404724, + 0.05252094194293022, + 0.05676804855465889 + ], + "credit_map": [ + 0.02552209049463272, + 0.020289218053221703, + 0.01789966970682144, + 0.011868107132613659, + 0.025202695280313492, + 0.020665239542722702, + 0.011208338662981987, + 0.014630298130214214, + 0.020244605839252472, + 0.012986126355826855, + 0.016349000856280327, + 0.012290561571717262, + 0.008719249628484249, + 0.009449541568756104, + 0.009127744473516941, + 0.013138750568032265, + 0.022572621703147888, + 0.016553815454244614, + 0.02094402350485325, + 0.024496858939528465, + 0.026788191869854927, + 0.02976386249065399, + 0.013537708669900894, + 0.011708761565387249, + 0.010790831409394741, + 0.006440056953579187, + 0.02488953433930874, + 0.0060737403109669685, + 0.031521450728178024, + 0.052505433559417725, + 0.007313757669180632, + 0.00711072888225317, + 0.007894054986536503, + 0.017546484246850014, + 0.006503459997475147, + 0.01700608618557453, + 0.008800237439572811, + 0.0072229471988976, + 0.009734602645039558, + 0.007619746960699558, + 0.0042293379083275795, + 0.006693514995276928, + 0.004909328185021877, + 0.00920676440000534, + 0.004308917559683323, + 0.003948355559259653, + 0.006857407279312611, + 0.0035854571033269167, + 0.005564604885876179, + 0.010180615819990635, + 0.004296650644391775, + 0.008365935645997524, + 0.008108877576887608, + 0.009752301499247551, + 0.01593416929244995, + 0.015284264460206032, + 0.01508260052651167, + 0.01452222652733326, + 0.027271080762147903, + 0.012437296099960804, + 0.013609128072857857, + 0.017385009676218033, + 0.010813293978571892, + 0.0134336082264781, + 0.004679088015109301, + 0.004233531653881073, + 0.007608249317854643, + 0.007287335116416216, + 0.00742219015955925, + 0.004016282968223095, + 0.008909646421670914, + 0.008364612236618996, + 0.010662386193871498, + 0.006313981954008341, + 0.006080235820263624, + 0.005656634457409382, + 0.007047839928418398, + 0.010130295529961586, + 0.005771251395344734, + 0.005105419550091028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.012565725483000278 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636165712028742 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.007636165712028742, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636165712028742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04716883599758148, + 0.05280774459242821, + 0.023057784885168076, + 0.4847770929336548, + 0.03788375481963158, + 0.04980764910578728, + 0.09034846723079681, + 0.08436329662799835, + 0.04070476070046425, + 0.0890805646777153 + ], + "credit_map": [ + 0.0063672843389213085, + 0.005461608991026878, + 0.0057815914042294025, + 0.00372371100820601, + 0.007775221951305866, + 0.007751645054668188, + 0.004230740014463663, + 0.006077032070606947, + 0.007697582244873047, + 0.007061190903186798, + 0.00661885179579258, + 0.006126497872173786, + 0.005286845378577709, + 0.007206210400909185, + 0.005598809104412794, + 0.007211757358163595, + 0.0031800298020243645, + 0.0025530625134706497, + 0.005095951724797487, + 0.0029957194346934557, + 0.002564098685979843, + 0.0034776064567267895, + 0.0015926903579384089, + 0.0015986260259523988, + 0.01955222338438034, + 0.009354384616017342, + 0.1539359986782074, + 0.013528247363865376, + 0.05757307633757591, + 0.19943080842494965, + 0.009867655113339424, + 0.021534740924835205, + 0.0044490923173725605, + 0.005293786991387606, + 0.0034984054509550333, + 0.0049823252484202385, + 0.005221051629632711, + 0.0035256799310445786, + 0.0053824083879590034, + 0.00553100323304534, + 0.005502461921423674, + 0.006667202338576317, + 0.009749048389494419, + 0.007391591556370258, + 0.0041377912275493145, + 0.004551809746772051, + 0.007449249736964703, + 0.004358493257313967, + 0.009467311203479767, + 0.008038891479372978, + 0.005125127732753754, + 0.006568815093487501, + 0.012605573982000351, + 0.014928345568478107, + 0.021604426205158234, + 0.012009975500404835, + 0.007353365886956453, + 0.009962836280465126, + 0.012954323552548885, + 0.010322598740458488, + 0.007947423495352268, + 0.015639115124940872, + 0.006349497940391302, + 0.01383413840085268, + 0.0024694446474313736, + 0.002648912835866213, + 0.00576620502397418, + 0.005417068488895893, + 0.007305314298719168, + 0.0039344606921076775, + 0.00766206718981266, + 0.005501284264028072, + 0.0063431900925934315, + 0.018254248425364494, + 0.005867792293429375, + 0.011149568483233452, + 0.014231543056666851, + 0.014096281491219997, + 0.01163889653980732, + 0.007499041501432657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.007636165712028742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291219025850296 + ], + "opposite_logprob_tokens": [ + -3.7732911109924316 + ], + "logprob_reward": -0.023291219025850296, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291219025850296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2531638741493225, + 0.1076602190732956, + 0.05182526260614395, + 0.1011182963848114, + 0.1156935840845108, + 0.031191125512123108, + 0.06727302819490433, + 0.07624933868646622, + 0.1495535671710968, + 0.04627170413732529 + ], + "credit_map": [ + 0.04917662963271141, + 0.0204947367310524, + 0.029001647606492043, + 0.01984698697924614, + 0.02982212044298649, + 0.033723361790180206, + 0.027314454317092896, + 0.04378392547369003, + 0.01995239034295082, + 0.009313729591667652, + 0.019799500703811646, + 0.013140142895281315, + 0.010932530276477337, + 0.012542147189378738, + 0.008718850091099739, + 0.01326091680675745, + 0.006866276264190674, + 0.005538524128496647, + 0.010926944203674793, + 0.007852871902287006, + 0.00537421740591526, + 0.007114895153790712, + 0.004406843334436417, + 0.0037446916103363037, + 0.006243812385946512, + 0.007914294488728046, + 0.021075047552585602, + 0.008784000761806965, + 0.01201093103736639, + 0.02967221289873123, + 0.008944857865571976, + 0.00647314777597785, + 0.014410571195185184, + 0.024258874356746674, + 0.012620980851352215, + 0.02447729930281639, + 0.010504964739084244, + 0.012958448380231857, + 0.008776772767305374, + 0.007685669232159853, + 0.004077036399394274, + 0.0029549605678766966, + 0.003736818442121148, + 0.0028484996873885393, + 0.004827885422855616, + 0.003452135482802987, + 0.0051672738045454025, + 0.004126516170799732, + 0.004141794051975012, + 0.006443595979362726, + 0.005386630073189735, + 0.00657446077093482, + 0.009596118703484535, + 0.01005288865417242, + 0.013579562306404114, + 0.011497977189719677, + 0.005691979080438614, + 0.005338112823665142, + 0.01323398482054472, + 0.006855680141597986, + 0.009218517690896988, + 0.012308160774409771, + 0.009326332248747349, + 0.01427657250314951, + 0.017065178602933884, + 0.01271668728441, + 0.024181772023439407, + 0.024263475090265274, + 0.018946459516882896, + 0.011970147490501404, + 0.013915044255554676, + 0.026494793593883514, + 0.006946817506104708, + 0.007536072283983231, + 0.006966533605009317, + 0.005654605571180582, + 0.005966372787952423, + 0.0035575595684349537, + 0.00606099097058177, + 0.0035827537067234516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.023291219025850296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot from a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the right as a worn bus with peeling paint lurches to a halt on the curb?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a weary woman in faded blue canvas overalls with black hair in a loose ponytail visible exiting the bus?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by dim interior bus light contrasting with cool dusk outside, featuring damp asphalt reflecting streetlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video convey a quiet atmosphere of fatigue with interweaving cold blue and warm yellow light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615036141127348 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003615036141127348, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615036141127348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16629783809185028, + 0.04344398155808449, + 0.06875202804803848, + 0.28531089425086975, + 0.05809846147894859, + 0.03660862520337105, + 0.12388300895690918, + 0.06338122487068176, + 0.10181131958961487, + 0.052412599325180054 + ], + "credit_map": [ + 0.02900264970958233, + 0.01957201585173607, + 0.018033809959888458, + 0.020051022991538048, + 0.019257284700870514, + 0.02245866134762764, + 0.01632501929998398, + 0.021597372367978096, + 0.00614609569311142, + 0.004665730521082878, + 0.008368217386305332, + 0.004832822363823652, + 0.003349987557157874, + 0.005579910706728697, + 0.002968955086544156, + 0.007532264571636915, + 0.008515959605574608, + 0.006924735847860575, + 0.012591320089995861, + 0.010146265849471092, + 0.008734794333577156, + 0.0074272602796554565, + 0.007478478364646435, + 0.006933216005563736, + 0.013797453604638577, + 0.0081474669277668, + 0.09109941124916077, + 0.009523632004857063, + 0.03199439495801926, + 0.11182518303394318, + 0.006555813364684582, + 0.012367542833089828, + 0.004809215199202299, + 0.01142144575715065, + 0.003552057547494769, + 0.00908732321113348, + 0.008326525799930096, + 0.008190339431166649, + 0.006037447135895491, + 0.006674109026789665, + 0.004397652577608824, + 0.003670594422146678, + 0.004744550678879023, + 0.004704356659203768, + 0.004836020525544882, + 0.004131329711526632, + 0.006438825279474258, + 0.003685293486341834, + 0.012178800068795681, + 0.013820791617035866, + 0.008210447616875172, + 0.011406157165765762, + 0.014593984931707382, + 0.014335380867123604, + 0.02701546624302864, + 0.022321976721286774, + 0.006667126435786486, + 0.007978617213666439, + 0.011872571893036366, + 0.005269214976578951, + 0.006784289143979549, + 0.011533033102750778, + 0.005080877803266048, + 0.008195496164262295, + 0.008653703145682812, + 0.007775652687996626, + 0.013069991022348404, + 0.014085507951676846, + 0.017138568684458733, + 0.008552107028663158, + 0.020127490162849426, + 0.012408298440277576, + 0.008044631220400333, + 0.0060519687831401825, + 0.003042470198124647, + 0.004011726938188076, + 0.007433719467371702, + 0.013926234096288681, + 0.005423135124146938, + 0.004478712100535631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.003615036141127348 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_1/index_tensor([8105, 8105])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_1/index_tensor([8105, 8105])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6922bbc6c8b4d648f0a2638262f370d0aee01914 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_1/index_tensor([8105, 8105])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, high-angle, overhead shot slowly pulls back to reveal a magnificent spectacle: a boundless sea of golden clouds roiling violently, like a seething ocean of molten gold. Vast waves collide, forming deep eddies and towering crests. At the center of this violent and sacred scene, the Monkey King, clad in a magnificent suit of armor, stands proud. He wears a phoenix-winged, purple-gold crown, its feathers trembling in the strong wind. He clad in golden chain mail, each plate reflecting the almighty golden light from the heavens, shimmering with brilliance. His feet, clad in lotus-silk, cloud-walking shoes, firmly anchor him above the turbulent sea of clouds. His powerful arm grips the heavy Ruyi Jingu Staff, one end of which is deeply embedded in the swirling clouds. Like a stabilizing force, his figure remains unshaken, regardless of the turbulent waves around him. His signature red cape dances wildly behind him, creating a striking contrast against the golden backdrop. Although it is viewed from a high altitude, one can still feel his unruly aura and his eyes as sharp as torches, as if they have penetrated through layers of clouds and mist, gazing into the unknown distance. The entire picture is filled with a mythological atmosphere of grandeur, solemnity and endless power.", + "seed": 7475, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_1/index_8105_seed_7475_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.018955570785328744, + "vlm_logprob": -0.018955570785328744, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.018955570785328744, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07052408158779144 + ], + "opposite_logprob_tokens": [ + -2.820523977279663 + ], + "logprob_reward": -0.07052408158779144, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07052408158779144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19290007650852203, + 0.0823267325758934, + 0.05811750143766403, + 0.0685947835445404, + 0.14761096239089966, + 0.04930279031395912, + 0.1619018018245697, + 0.05881301313638687, + 0.0722222849726677, + 0.10820996761322021 + ], + "credit_map": [ + 0.0248800627887249, + 0.024390678852796555, + 0.024961154907941818, + 0.015491734258830547, + 0.015892039984464645, + 0.039848681539297104, + 0.023185336962342262, + 0.02425040304660797, + 0.008803343400359154, + 0.009103070013225079, + 0.01025347039103508, + 0.011214000172913074, + 0.008777140639722347, + 0.011300441808998585, + 0.009263675659894943, + 0.013611597009003162, + 0.011403701268136501, + 0.00871520396322012, + 0.006630637217313051, + 0.004915402736514807, + 0.005337190348654985, + 0.005828759167343378, + 0.005964980460703373, + 0.009321630001068115, + 0.020400309935212135, + 0.02068275958299637, + 0.005862889811396599, + 0.0037719665560871363, + 0.004698204807937145, + 0.0027668981347233057, + 0.003434052225202322, + 0.006977711338549852, + 0.01691160537302494, + 0.017900243401527405, + 0.02151833288371563, + 0.022022772580385208, + 0.012082548812031746, + 0.026257960125803947, + 0.012109540402889252, + 0.01880798116326332, + 0.0065843649208545685, + 0.008143826387822628, + 0.00825718604028225, + 0.005389786325395107, + 0.004160087089985609, + 0.005903720390051603, + 0.005008789245039225, + 0.0058550359681248665, + 0.02328047715127468, + 0.0369771383702755, + 0.01453656330704689, + 0.015529053285717964, + 0.010677820071578026, + 0.01351583655923605, + 0.00913766585290432, + 0.03824726864695549, + 0.009866761043667793, + 0.006578534375876188, + 0.011848093010485172, + 0.006769997533410788, + 0.003467364702373743, + 0.00683108763769269, + 0.0037405677139759064, + 0.009710608050227165, + 0.00903968047350645, + 0.00881151482462883, + 0.008535144850611687, + 0.008844785392284393, + 0.009568169713020325, + 0.009044738486409187, + 0.009128939360380173, + 0.00924932025372982, + 0.01176175195723772, + 0.008936251513659954, + 0.014480368234217167, + 0.011135734617710114, + 0.017745884135365486, + 0.012614068575203419, + 0.017836004495620728, + 0.013699909672141075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.07052408158779144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036353471223264933 + ], + "opposite_logprob_tokens": [ + -5.628635406494141 + ], + "logprob_reward": -0.0036353471223264933, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036353471223264933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21511560678482056, + 0.16593000292778015, + 0.05247129499912262, + 0.07184004783630371, + 0.1631074696779251, + 0.06556736677885056, + 0.052504751831293106, + 0.03961993008852005, + 0.10187853872776031, + 0.07196501642465591 + ], + "credit_map": [ + 0.028342051431536674, + 0.02594047226011753, + 0.027879023924469948, + 0.01960853487253189, + 0.01857011578977108, + 0.029454123228788376, + 0.03190441057085991, + 0.03341686725616455, + 0.023043682798743248, + 0.019787196069955826, + 0.02388111874461174, + 0.02203403227031231, + 0.015481045469641685, + 0.019352931529283524, + 0.017521932721138, + 0.02482806146144867, + 0.006128314416855574, + 0.0075761741027235985, + 0.008127504028379917, + 0.005618835333734751, + 0.006615100894123316, + 0.005322375800460577, + 0.006159007549285889, + 0.006923980545252562, + 0.008631065487861633, + 0.010748447850346565, + 0.015814602375030518, + 0.007519489619880915, + 0.004480829928070307, + 0.008084481582045555, + 0.0049103605560958385, + 0.011650767177343369, + 0.020585352554917336, + 0.020298542454838753, + 0.023915506899356842, + 0.025344764813780785, + 0.012353337369859219, + 0.02383662946522236, + 0.013641794212162495, + 0.023131536319851875, + 0.007364256307482719, + 0.010749677196145058, + 0.011136348359286785, + 0.012031199410557747, + 0.004520520102232695, + 0.008036255836486816, + 0.004940798506140709, + 0.006788311991840601, + 0.0064385575242340565, + 0.0061667440459132195, + 0.005757494363933802, + 0.006224488839507103, + 0.006605620961636305, + 0.005256658419966698, + 0.008780893869698048, + 0.0072742970660328865, + 0.005978014320135117, + 0.003714389633387327, + 0.006818085443228483, + 0.0058951787650585175, + 0.003608910134062171, + 0.003927929326891899, + 0.0041406224481761456, + 0.00553679745644331, + 0.015479318797588348, + 0.010377665050327778, + 0.014564157463610172, + 0.01118288841098547, + 0.01191968098282814, + 0.011982455849647522, + 0.011333934031426907, + 0.015038444660604, + 0.009449551813304424, + 0.009539859369397163, + 0.012531151995062828, + 0.0053195166401565075, + 0.007268351502716541, + 0.008322461508214474, + 0.00849865097552538, + 0.01103547029197216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0036353471223264933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738083902746439 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738083902746439, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738083902746439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12368708848953247, + 0.08710544556379318, + 0.11736981570720673, + 0.13342304527759552, + 0.20029202103614807, + 0.08013566583395004, + 0.05819329619407654, + 0.08026014268398285, + 0.06255228817462921, + 0.056981172412633896 + ], + "credit_map": [ + 0.013708030804991722, + 0.014874192886054516, + 0.01837940141558647, + 0.010683652944862843, + 0.016915231943130493, + 0.018067965283989906, + 0.016188735142350197, + 0.014869877137243748, + 0.012647012248635292, + 0.01451778318732977, + 0.011446010321378708, + 0.007461903151124716, + 0.010328845120966434, + 0.008386794477701187, + 0.008564535528421402, + 0.013752559199929237, + 0.01147842314094305, + 0.011117455549538136, + 0.020135750994086266, + 0.011053916998207569, + 0.013809184543788433, + 0.014454022981226444, + 0.014753005467355251, + 0.020568057894706726, + 0.03224802762269974, + 0.036080501973629, + 0.011360161006450653, + 0.007811106275767088, + 0.009719658643007278, + 0.007619089912623167, + 0.01070019043982029, + 0.01788431406021118, + 0.028575610369443893, + 0.023995844647288322, + 0.024976832792162895, + 0.02472272701561451, + 0.024004796519875526, + 0.027158381417393684, + 0.025768209248781204, + 0.02108961157500744, + 0.008512760512530804, + 0.007981067523360252, + 0.009747662581503391, + 0.009045540355145931, + 0.012222670949995518, + 0.011655259877443314, + 0.012209092266857624, + 0.008761616423726082, + 0.007877473719418049, + 0.010460805147886276, + 0.0068923053331673145, + 0.004362208768725395, + 0.005203058011829853, + 0.005999348592013121, + 0.0063922712579369545, + 0.011005821637809277, + 0.010466295294463634, + 0.006911055184900761, + 0.013691567815840244, + 0.009359413757920265, + 0.009857832454144955, + 0.007564508821815252, + 0.010866828262805939, + 0.01154264621436596, + 0.006672352086752653, + 0.006649401504546404, + 0.011525831185281277, + 0.005838973913341761, + 0.006038444582372904, + 0.0075468821451067924, + 0.004517360124737024, + 0.013763046823441982, + 0.0044251130893826485, + 0.0038046364206820726, + 0.011376108974218369, + 0.005574709735810757, + 0.007136679720133543, + 0.0072679040022194386, + 0.006763783749192953, + 0.010632235556840897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.006738083902746439 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783296845853329 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783296845853329, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783296845853329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15674586594104767, + 0.0662412941455841, + 0.1024068146944046, + 0.08326825499534607, + 0.09800990670919418, + 0.1299310326576233, + 0.06491395831108093, + 0.16208140552043915, + 0.07878699153661728, + 0.057614509016275406 + ], + "credit_map": [ + 0.021458551287651062, + 0.017729509621858597, + 0.022009341046214104, + 0.015615232288837433, + 0.016584742814302444, + 0.030202573165297508, + 0.014329779893159866, + 0.018816135823726654, + 0.012333988212049007, + 0.010336116887629032, + 0.0072205932810902596, + 0.0067133731208741665, + 0.007329388987272978, + 0.006610603537410498, + 0.007090242579579353, + 0.008606990799307823, + 0.01744365133345127, + 0.0111036142334342, + 0.012957656756043434, + 0.006865995470434427, + 0.007908566854894161, + 0.007319633848965168, + 0.007827060297131538, + 0.030980635434389114, + 0.017963139340281487, + 0.02270699106156826, + 0.008030525408685207, + 0.005534699652343988, + 0.006967613473534584, + 0.005333591252565384, + 0.0071808332577347755, + 0.00955086387693882, + 0.011404009535908699, + 0.012798286974430084, + 0.01185443066060543, + 0.010550973936915398, + 0.014301266521215439, + 0.012942309491336346, + 0.014531837776303291, + 0.009626789949834347, + 0.016169697046279907, + 0.029125988483428955, + 0.016809366643428802, + 0.0133003955706954, + 0.01474467571824789, + 0.014802662655711174, + 0.012163996696472168, + 0.012814247980713844, + 0.008999387733638287, + 0.009005546569824219, + 0.00939639937132597, + 0.007705199997872114, + 0.006894655060023069, + 0.007833492942154408, + 0.007460708264261484, + 0.007618570700287819, + 0.02144087292253971, + 0.016755657270550728, + 0.03269396722316742, + 0.020018478855490685, + 0.01239071600139141, + 0.017886508256196976, + 0.015570208430290222, + 0.025324994698166847, + 0.008237863890826702, + 0.0074777607806026936, + 0.011728084646165371, + 0.010282758623361588, + 0.007206805516034365, + 0.010505028069019318, + 0.009502398781478405, + 0.013846294023096561, + 0.006111708004027605, + 0.005578318610787392, + 0.011361557058990002, + 0.004319310188293457, + 0.00677257077768445, + 0.006368318106979132, + 0.006905938033014536, + 0.010196785442531109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.009783296845853329 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097044467926025 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.004097044467926025, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097044467926025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13950000703334808, + 0.044547390192747116, + 0.10371517390012741, + 0.14367826282978058, + 0.0671640932559967, + 0.08389125764369965, + 0.05173517018556595, + 0.06949874758720398, + 0.143739715218544, + 0.152530238032341 + ], + "credit_map": [ + 0.014815899543464184, + 0.013683163560926914, + 0.018924254924058914, + 0.013901433907449245, + 0.016877971589565277, + 0.016515236347913742, + 0.023689040914177895, + 0.02109300158917904, + 0.006286827847361565, + 0.004642142448574305, + 0.005872098263353109, + 0.004496449138969183, + 0.005666943266987801, + 0.004287794232368469, + 0.005790154915302992, + 0.007504977285861969, + 0.028216123580932617, + 0.00918202381581068, + 0.009753852151334286, + 0.006939209997653961, + 0.006731000728905201, + 0.010950230062007904, + 0.007969276048243046, + 0.02397346869111061, + 0.034639738500118256, + 0.04129405319690704, + 0.011275121010839939, + 0.008805242367088795, + 0.009757189080119133, + 0.007857459597289562, + 0.011382774449884892, + 0.018666677176952362, + 0.009420357644557953, + 0.00858944933861494, + 0.007736464962363243, + 0.009433670900762081, + 0.008919321931898594, + 0.0074741230346262455, + 0.008319801650941372, + 0.0072709000669419765, + 0.011886855587363243, + 0.010896771214902401, + 0.012510267086327076, + 0.01073099672794342, + 0.009906856343150139, + 0.008587176911532879, + 0.010533166117966175, + 0.008839164860546589, + 0.008675755001604557, + 0.01064083818346262, + 0.004380226135253906, + 0.004996557254344225, + 0.004852959420531988, + 0.004271751269698143, + 0.005279346834868193, + 0.00863773562014103, + 0.007165617309510708, + 0.006681682541966438, + 0.013474681414663792, + 0.009924770332872868, + 0.007759811356663704, + 0.007424777373671532, + 0.007774306461215019, + 0.009293101727962494, + 0.018803587183356285, + 0.014405832625925541, + 0.01919490471482277, + 0.014054317027330399, + 0.018328431993722916, + 0.017645400017499924, + 0.01910441927611828, + 0.02220282331109047, + 0.01417851634323597, + 0.015845684334635735, + 0.02213793806731701, + 0.01230537611991167, + 0.021811820566654205, + 0.02049436792731285, + 0.02717476896941662, + 0.01858176663517952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.004097044467926025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, high-angle, overhead shot slowly pulls back to reveal a magnificent spectacle: a boundless sea of golden clouds roiling violently, like a seething ocean of molten gold. Vast waves collide, forming deep eddies and towering crests. At the center of this violent and sacred scene, the Monkey King, clad in a magnificent suit of armor, stands proud. He wears a phoenix-winged, purple-gold crown, its feathers trembling in the strong wind. He clad in golden chain mail, each plate reflecting the almighty golden light from the heavens, shimmering with brilliance. His feet, clad in lotus-silk, cloud-walking shoes, firmly anchor him above the turbulent sea of clouds. His powerful arm grips the heavy Ruyi Jingu Staff, one end of which is deeply embedded in the swirling clouds. Like a stabilizing force, his figure remains unshaken, regardless of the turbulent waves around him. His signature red cape dances wildly behind him, creating a striking contrast against the golden backdrop. Although it is viewed from a high altitude, one can still feel his unruly aura and his eyes as sharp as torches, as if they have penetrated through layers of clouds and mist, gazing into the unknown distance. The entire picture is filled with a mythological atmosphere of grandeur, solemnity and endless power.", + "seed": 7476, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_1/index_8105_seed_7476_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0776233988814056, + "vlm_logprob": -0.0776233988814056, + "vlm_yes_no_margin": 4.374999988079071, + "vlm_token_credit": -0.0776233988814056, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3594258427619934 + ], + "opposite_logprob_tokens": [ + -1.2344257831573486 + ], + "logprob_reward": -0.3594258427619934, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3594258427619934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22208626568317413, + 0.07447631657123566, + 0.21367591619491577, + 0.09963297843933105, + 0.03570206090807915, + 0.06154125556349754, + 0.06931092590093613, + 0.03383095562458038, + 0.14172369241714478, + 0.04801966995000839 + ], + "credit_map": [ + 0.022061405703425407, + 0.023893380537629128, + 0.0350799523293972, + 0.01795700006186962, + 0.019834216684103012, + 0.03274548426270485, + 0.01833237148821354, + 0.05218241736292839, + 0.00968928448855877, + 0.009425113908946514, + 0.007318479474633932, + 0.0065541937947273254, + 0.00859044399112463, + 0.009201504290103912, + 0.00998204667121172, + 0.013715239241719246, + 0.013818282634019852, + 0.009022042155265808, + 0.07189807295799255, + 0.013798254542052746, + 0.010219903662800789, + 0.06624733656644821, + 0.020892295986413956, + 0.007779715117067099, + 0.01823178492486477, + 0.0070859938859939575, + 0.015606356784701347, + 0.008086210116744041, + 0.009736400097608566, + 0.016387954354286194, + 0.010426266118884087, + 0.01407200563699007, + 0.005243966355919838, + 0.0034038033336400986, + 0.004750496707856655, + 0.004318352323025465, + 0.005305870436131954, + 0.004410632885992527, + 0.0034537920728325844, + 0.004815143998712301, + 0.011120178736746311, + 0.004829066805541515, + 0.00786218885332346, + 0.0064728474244475365, + 0.004731283988803625, + 0.012037708424031734, + 0.006612368393689394, + 0.007875609211623669, + 0.011987068690359592, + 0.00653166463598609, + 0.010290585458278656, + 0.006205373909324408, + 0.01017188560217619, + 0.011264362372457981, + 0.005841691978275776, + 0.00701828533783555, + 0.005488615483045578, + 0.003523297840729356, + 0.004845426883548498, + 0.003915719222277403, + 0.003690562443807721, + 0.005005513317883015, + 0.0036061627324670553, + 0.003755655139684677, + 0.025416284799575806, + 0.01772088184952736, + 0.018289262428879738, + 0.015035361051559448, + 0.013760751113295555, + 0.02102692425251007, + 0.015805810689926147, + 0.01466839388012886, + 0.004982071463018656, + 0.005220526363700628, + 0.007901014760136604, + 0.00654183654114604, + 0.005313357803970575, + 0.006136947777122259, + 0.004838889464735985, + 0.007085022982209921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.3594258427619934 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025095418095588684 + ], + "opposite_logprob_tokens": [ + -6.002509593963623 + ], + "logprob_reward": -0.0025095418095588684, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025095418095588684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2769908607006073, + 0.07822544872760773, + 0.2918056845664978, + 0.17395226657390594, + 0.021289143711328506, + 0.01923874020576477, + 0.0471019446849823, + 0.02222687192261219, + 0.03515299782156944, + 0.03401600942015648 + ], + "credit_map": [ + 0.029086964204907417, + 0.022275863215327263, + 0.060177914798259735, + 0.01714823767542839, + 0.018960746005177498, + 0.031116874888539314, + 0.017717450857162476, + 0.0805068165063858, + 0.013028531335294247, + 0.007342047058045864, + 0.010114351287484169, + 0.005018356256186962, + 0.00845066737383604, + 0.01008345652371645, + 0.0073211826384067535, + 0.016866864636540413, + 0.018717292696237564, + 0.009733719751238823, + 0.11407943069934845, + 0.009697631001472473, + 0.011124076321721077, + 0.10633055865764618, + 0.011771309189498425, + 0.010351651348173618, + 0.030326003208756447, + 0.013945095241069794, + 0.02372361347079277, + 0.015987198799848557, + 0.015472626313567162, + 0.03281195089221001, + 0.01310737058520317, + 0.028578413650393486, + 0.0037494832649827003, + 0.0019776024855673313, + 0.0021104065235704184, + 0.003215606790035963, + 0.002769958693534136, + 0.0029655422549694777, + 0.00210370821878314, + 0.002396834548562765, + 0.003998824395239353, + 0.0019598612561821938, + 0.002435440430417657, + 0.0015011673094704747, + 0.0021783492993563414, + 0.0028913605492562056, + 0.0014861433301120996, + 0.00278759328648448, + 0.010261684656143188, + 0.004709768109023571, + 0.006965137552469969, + 0.004893395584076643, + 0.0053281644359230995, + 0.007377850357443094, + 0.0034181326627731323, + 0.004147812258452177, + 0.004057938698679209, + 0.0029565291479229927, + 0.003381675109267235, + 0.0019510275451466441, + 0.0026406242977827787, + 0.0027756141498684883, + 0.0022696740925312042, + 0.0021937887649983168, + 0.006366835441440344, + 0.004059408791363239, + 0.005204024724662304, + 0.004255951847881079, + 0.00310239614918828, + 0.005149677395820618, + 0.0037914803251624107, + 0.003223221283406019, + 0.00475096283480525, + 0.003563065081834793, + 0.0065574850887060165, + 0.003726473543792963, + 0.0025125443935394287, + 0.003910984378308058, + 0.002950166119262576, + 0.006044328212738037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0025095418095588684 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978388637304306 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.00978388637304306, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978388637304306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22394217550754547, + 0.06516934186220169, + 0.44967713952064514, + 0.05118950083851814, + 0.0201119352132082, + 0.021363988518714905, + 0.07335905730724335, + 0.03670496866106987, + 0.03266315162181854, + 0.025818733498454094 + ], + "credit_map": [ + 0.022450527176260948, + 0.02417750656604767, + 0.03372446447610855, + 0.01785343699157238, + 0.03397860378026962, + 0.02370443008840084, + 0.024451687932014465, + 0.04360151290893555, + 0.009400656446814537, + 0.006658457685261965, + 0.00643745856359601, + 0.007162213325500488, + 0.009139640256762505, + 0.00976102240383625, + 0.008097951300442219, + 0.008511943742632866, + 0.028270822018384933, + 0.01241039764136076, + 0.17364661395549774, + 0.01393347978591919, + 0.016516033560037613, + 0.17323514819145203, + 0.01717858389019966, + 0.01448605116456747, + 0.006367577705532312, + 0.004960968159139156, + 0.005344684701412916, + 0.004515521693974733, + 0.007005263585597277, + 0.008197649382054806, + 0.005689283367246389, + 0.00910855084657669, + 0.002675477182492614, + 0.0023946354631334543, + 0.0027032336220145226, + 0.002404827391728759, + 0.0027715545147657394, + 0.0023093121126294136, + 0.002614395460113883, + 0.002238499466329813, + 0.002258995082229376, + 0.0031442486215382814, + 0.00246525090187788, + 0.0020845711696892977, + 0.004166022874414921, + 0.0026657867711037397, + 0.0027399065438658, + 0.0018392057390883565, + 0.02075807750225067, + 0.008667130954563618, + 0.012050983496010303, + 0.006592682562768459, + 0.005086811259388924, + 0.008181489072740078, + 0.006223202217370272, + 0.005798683036118746, + 0.006338163744658232, + 0.004662555642426014, + 0.0033818925730884075, + 0.0032150305341929197, + 0.005656078457832336, + 0.004579171072691679, + 0.0038945977576076984, + 0.004977475851774216, + 0.005599547177553177, + 0.0037873233668506145, + 0.004707058891654015, + 0.00382805778644979, + 0.00380186946131289, + 0.00444524921476841, + 0.0031453906558454037, + 0.0033486541360616684, + 0.004390612710267305, + 0.003120927605777979, + 0.002927979454398155, + 0.002532192738726735, + 0.0030057700350880623, + 0.00304601714015007, + 0.0032619773410260677, + 0.0035332569386810064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.00978388637304306 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419894304126501 + ], + "opposite_logprob_tokens": [ + -4.2641987800598145 + ], + "logprob_reward": -0.01419894304126501, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419894304126501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3756423592567444, + 0.06408348679542542, + 0.28294840455055237, + 0.04647667333483696, + 0.05832988768815994, + 0.018438922241330147, + 0.031152866780757904, + 0.023075824603438377, + 0.05864972621202469, + 0.041201889514923096 + ], + "credit_map": [ + 0.03449082747101784, + 0.019267462193965912, + 0.10189562290906906, + 0.019760245457291603, + 0.020445501431822777, + 0.025688277557492256, + 0.020891230553388596, + 0.13320320844650269, + 0.00809912383556366, + 0.008105368353426456, + 0.005454178433865309, + 0.011774430051445961, + 0.0056474339216947556, + 0.008201667107641697, + 0.009402984753251076, + 0.0073983012698590755, + 0.01752481982111931, + 0.011598042212426662, + 0.10712677240371704, + 0.010388600639998913, + 0.011477912776172161, + 0.1011100560426712, + 0.013950850814580917, + 0.009771350771188736, + 0.007326842285692692, + 0.003066179808229208, + 0.005978057626634836, + 0.004711978603154421, + 0.0040476033464074135, + 0.009162947535514832, + 0.0044588856399059296, + 0.00772417988628149, + 0.0078042722307145596, + 0.006536779925227165, + 0.006193988490849733, + 0.009252574294805527, + 0.009493988938629627, + 0.006205069366842508, + 0.008659640327095985, + 0.004183582495898008, + 0.0033270521089434624, + 0.0018732526805251837, + 0.0025816410779953003, + 0.001535717281512916, + 0.0019677800592035055, + 0.0027750246226787567, + 0.002107214415445924, + 0.002271241508424282, + 0.0049412124790251255, + 0.0037084450013935566, + 0.004212814383208752, + 0.0027579222805798054, + 0.004453228320926428, + 0.004638247657567263, + 0.002797250170260668, + 0.0036437460221350193, + 0.0036362011451274157, + 0.00196242006495595, + 0.0022268088068813086, + 0.0020223113242536783, + 0.004639915190637112, + 0.0027418923564255238, + 0.002865833230316639, + 0.0029804466757923365, + 0.010883443057537079, + 0.005876387003809214, + 0.01040117908269167, + 0.006760759744793177, + 0.005269707180559635, + 0.008616109378635883, + 0.005074504297226667, + 0.005767642054706812, + 0.00551244243979454, + 0.004748353734612465, + 0.006464928388595581, + 0.0038929933216422796, + 0.005729238968342543, + 0.004706368315964937, + 0.004505130462348461, + 0.005642435979098082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.01419894304126501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002198780421167612 + ], + "opposite_logprob_tokens": [ + -6.127198696136475 + ], + "logprob_reward": -0.002198780421167612, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002198780421167612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39703646302223206, + 0.05231080204248428, + 0.2587541341781616, + 0.0581403486430645, + 0.03125867247581482, + 0.018290113657712936, + 0.030149176716804504, + 0.04814334958791733, + 0.03541216626763344, + 0.07050470262765884 + ], + "credit_map": [ + 0.03403334692120552, + 0.02535933069884777, + 0.09527727216482162, + 0.02561747282743454, + 0.03814392164349556, + 0.0289839468896389, + 0.03436073660850525, + 0.11526046693325043, + 0.006383446976542473, + 0.007236274424940348, + 0.004801397677510977, + 0.008280619978904724, + 0.0056123072281479836, + 0.0061690290458500385, + 0.00809136126190424, + 0.005736364517360926, + 0.015150361694395542, + 0.011184087954461575, + 0.0883215144276619, + 0.012229733169078827, + 0.011890226975083351, + 0.08991046249866486, + 0.020945174619555473, + 0.009122589603066444, + 0.009212289936840534, + 0.0043676369823515415, + 0.007864025421440601, + 0.004649808164685965, + 0.005856259260326624, + 0.00940100196748972, + 0.006019137799739838, + 0.010770192369818687, + 0.004317284096032381, + 0.0030547366477549076, + 0.003856328781694174, + 0.002358357422053814, + 0.005998699925839901, + 0.0028673922643065453, + 0.004361429251730442, + 0.004444447811692953, + 0.002700316719710827, + 0.0023053192999213934, + 0.002146846614778042, + 0.0017381436191499233, + 0.0025431697722524405, + 0.0026806804817169905, + 0.002181997522711754, + 0.0019936414901167154, + 0.003140389919281006, + 0.004270641598850489, + 0.002722420496866107, + 0.0027025765739381313, + 0.006541247479617596, + 0.0038818169850856066, + 0.003929192665964365, + 0.002960891928523779, + 0.006353739183396101, + 0.005954276770353317, + 0.004715658724308014, + 0.004444904625415802, + 0.009626481682062149, + 0.004971046932041645, + 0.007006474304944277, + 0.005070773418992758, + 0.005851569585502148, + 0.0042429640889167786, + 0.005165328737348318, + 0.004298903048038483, + 0.0037009860388934612, + 0.004805056843906641, + 0.004106718115508556, + 0.0032406412065029144, + 0.00962347723543644, + 0.005667094606906176, + 0.014305489137768745, + 0.0076723843812942505, + 0.004168772604316473, + 0.0068437629379332066, + 0.0066703492775559425, + 0.015553375706076622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.002198780421167612 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, high-angle, overhead shot slowly pulls back to reveal a magnificent spectacle: a boundless sea of golden clouds roiling violently, like a seething ocean of molten gold. Vast waves collide, forming deep eddies and towering crests. At the center of this violent and sacred scene, the Monkey King, clad in a magnificent suit of armor, stands proud. He wears a phoenix-winged, purple-gold crown, its feathers trembling in the strong wind. He clad in golden chain mail, each plate reflecting the almighty golden light from the heavens, shimmering with brilliance. His feet, clad in lotus-silk, cloud-walking shoes, firmly anchor him above the turbulent sea of clouds. His powerful arm grips the heavy Ruyi Jingu Staff, one end of which is deeply embedded in the swirling clouds. Like a stabilizing force, his figure remains unshaken, regardless of the turbulent waves around him. His signature red cape dances wildly behind him, creating a striking contrast against the golden backdrop. Although it is viewed from a high altitude, one can still feel his unruly aura and his eyes as sharp as torches, as if they have penetrated through layers of clouds and mist, gazing into the unknown distance. The entire picture is filled with a mythological atmosphere of grandeur, solemnity and endless power.", + "seed": 7477, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_1/index_8105_seed_7477_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07708004410378635, + "vlm_logprob": -0.07708004410378635, + "vlm_yes_no_margin": 4.399999988079071, + "vlm_token_credit": -0.07708004410378635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35837990045547485 + ], + "opposite_logprob_tokens": [ + -1.23337984085083 + ], + "logprob_reward": -0.35837990045547485, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.35837990045547485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17560423910617828, + 0.03425804525613785, + 0.027388066053390503, + 0.028741490095853806, + 0.023373132571578026, + 0.014417310245335102, + 0.5958548784255981, + 0.016697358340024948, + 0.044028833508491516, + 0.03963671252131462 + ], + "credit_map": [ + 0.015597439371049404, + 0.008965796791017056, + 0.06725116819143295, + 0.018292631953954697, + 0.006557611282914877, + 0.008982501924037933, + 0.010116441175341606, + 0.039840638637542725, + 0.004841603338718414, + 0.00340163242071867, + 0.007208171300590038, + 0.006690937094390392, + 0.0028146810363978148, + 0.003147489158436656, + 0.0037110077682882547, + 0.0024425263982266188, + 0.0027605798095464706, + 0.002541933674365282, + 0.009291812777519226, + 0.0031331023201346397, + 0.0014656545827165246, + 0.0022404391784220934, + 0.0019772390369325876, + 0.003977305255830288, + 0.003932035528123379, + 0.004613820463418961, + 0.0036555372644215822, + 0.004806204698979855, + 0.0036035238299518824, + 0.003334633307531476, + 0.0020689184311777353, + 0.0027268172707408667, + 0.00411571329459548, + 0.004420784767717123, + 0.0018960952293127775, + 0.001814733142964542, + 0.003977647516876459, + 0.0029164536390453577, + 0.002536657266318798, + 0.0016950464341789484, + 0.0019241441041231155, + 0.002196370856836438, + 0.0016210946487262845, + 0.001321193645708263, + 0.002632358344271779, + 0.0019926459062844515, + 0.0013037660392001271, + 0.001425736933015287, + 0.2387697845697403, + 0.09423752129077911, + 0.18992440402507782, + 0.023903414607048035, + 0.012768792919814587, + 0.014905060641467571, + 0.006454117596149445, + 0.014891787432134151, + 0.0016030597034841776, + 0.0022424000781029463, + 0.0020855178590863943, + 0.002595613244920969, + 0.0028685692232102156, + 0.0021404707804322243, + 0.001585505553521216, + 0.0015762208495289087, + 0.0032305798958986998, + 0.004047930706292391, + 0.006862233392894268, + 0.013652060180902481, + 0.004906201269477606, + 0.0024829735048115253, + 0.0033866397570818663, + 0.005460209678858519, + 0.009791091084480286, + 0.0058097317814826965, + 0.005371570587158203, + 0.0032836643513292074, + 0.0038615327794104815, + 0.004194523673504591, + 0.002903449349105358, + 0.004421147517859936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.35837990045547485 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002214005682617426 + ], + "opposite_logprob_tokens": [ + -6.127213954925537 + ], + "logprob_reward": -0.002214005682617426, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002214005682617426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1680518537759781, + 0.05203884840011597, + 0.0419653095304966, + 0.04031633213162422, + 0.04103567823767662, + 0.018959036096930504, + 0.5560451149940491, + 0.019784441217780113, + 0.03772391751408577, + 0.024079473689198494 + ], + "credit_map": [ + 0.01374832447618246, + 0.010622208938002586, + 0.05248475819826126, + 0.015162982977926731, + 0.011033411137759686, + 0.011251515708863735, + 0.014358781278133392, + 0.03938985615968704, + 0.006186212878674269, + 0.00446229986846447, + 0.006767997983843088, + 0.007939169183373451, + 0.0065501038916409016, + 0.009401997551321983, + 0.00706705404445529, + 0.0036640146281570196, + 0.004059633240103722, + 0.003611981403082609, + 0.011676345020532608, + 0.005384271498769522, + 0.004072302021086216, + 0.004564009606838226, + 0.003390737110748887, + 0.005206026136875153, + 0.007161384914070368, + 0.0063677262514829636, + 0.005829373374581337, + 0.004644013475626707, + 0.0034779298584908247, + 0.004867481533437967, + 0.004029463045299053, + 0.003938957117497921, + 0.007045013830065727, + 0.007877983152866364, + 0.004020256921648979, + 0.004569963086396456, + 0.004114070441573858, + 0.006893568206578493, + 0.0031396946869790554, + 0.003375125816091895, + 0.0020575018133968115, + 0.0020700471941381693, + 0.002818465931341052, + 0.0018130475655198097, + 0.003059887560084462, + 0.002780141541734338, + 0.0023192421067506075, + 0.00204070215113461, + 0.2242123931646347, + 0.11874386668205261, + 0.13969384133815765, + 0.022529689595103264, + 0.012316688895225525, + 0.017394281923770905, + 0.007076612673699856, + 0.014077777974307537, + 0.002199436305090785, + 0.0032036968041211367, + 0.0028491890989243984, + 0.0026804404333233833, + 0.0026459014043211937, + 0.0020797003526240587, + 0.0014828256098553538, + 0.0026432506274431944, + 0.004063898231834173, + 0.004721838515251875, + 0.005550788249820471, + 0.008445525541901588, + 0.004025237634778023, + 0.002674649702385068, + 0.0026494089979678392, + 0.005592568311840296, + 0.0052535077556967735, + 0.003669653320685029, + 0.0029055909253656864, + 0.0027482614386826754, + 0.002391352318227291, + 0.002048719674348831, + 0.0023839736822992563, + 0.00267841387540102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.002214005682617426 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536883354187012 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.012536883354187012, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536883354187012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22889098525047302, + 0.026998141780495644, + 0.0607234351336956, + 0.037542469799518585, + 0.051724545657634735, + 0.04341820999979973, + 0.4528151750564575, + 0.057722777128219604, + 0.021708911284804344, + 0.018455425277352333 + ], + "credit_map": [ + 0.022109000012278557, + 0.012659221887588501, + 0.09128823131322861, + 0.02196511998772621, + 0.00803615152835846, + 0.009820609353482723, + 0.012655524536967278, + 0.05035709962248802, + 0.0027377475053071976, + 0.002230040729045868, + 0.003999647218734026, + 0.0038153426721692085, + 0.003907320089638233, + 0.004753644112497568, + 0.0032114535570144653, + 0.0023429447319358587, + 0.005504063796252012, + 0.004829028155654669, + 0.018685538321733475, + 0.006096274126321077, + 0.006077471189200878, + 0.006513029336929321, + 0.005429461132735014, + 0.007588566746562719, + 0.0052759223617613316, + 0.004531598649919033, + 0.004404507111757994, + 0.005479760002344847, + 0.006874704733490944, + 0.004999788478016853, + 0.0029289661906659603, + 0.0030472204089164734, + 0.004561816807836294, + 0.00693809287622571, + 0.004740207921713591, + 0.004680347628891468, + 0.007857183925807476, + 0.011382101103663445, + 0.006305352319031954, + 0.005259442608803511, + 0.00427429610863328, + 0.006300807930529118, + 0.004571312107145786, + 0.005065298173576593, + 0.0074098194018006325, + 0.00709598558023572, + 0.004748173523694277, + 0.003952517174184322, + 0.17885927855968475, + 0.06607390940189362, + 0.14755800366401672, + 0.019540712237358093, + 0.008872460573911667, + 0.012234656140208244, + 0.006090682465583086, + 0.013585411012172699, + 0.006838048808276653, + 0.011882954277098179, + 0.007021917030215263, + 0.006121553480625153, + 0.008652836084365845, + 0.0069865696132183075, + 0.004758456721901894, + 0.005460440646857023, + 0.0018417572136968374, + 0.0030615893192589283, + 0.0026655623223632574, + 0.0037070359103381634, + 0.0031259346287697554, + 0.0018424824811518192, + 0.0024901358410716057, + 0.0029744133353233337, + 0.0027858775574713945, + 0.0022559240460395813, + 0.0016102987574413419, + 0.0014829920837655663, + 0.003875185502693057, + 0.0017812083242461085, + 0.0025476855225861073, + 0.002116253599524498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.012536883354187012 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780817665159702 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780817665159702, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780817665159702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06681124120950699, + 0.02649722993373871, + 0.0592542365193367, + 0.047161996364593506, + 0.034835562109947205, + 0.02335965633392334, + 0.630559504032135, + 0.0428205281496048, + 0.04219943657517433, + 0.026500554755330086 + ], + "credit_map": [ + 0.006802502553910017, + 0.01114298589527607, + 0.011456246487796307, + 0.00760142644867301, + 0.006281023845076561, + 0.007534629665315151, + 0.006862442474812269, + 0.00912998802959919, + 0.0021967270877212286, + 0.0022215344943106174, + 0.004061783663928509, + 0.005122814793139696, + 0.0032694374676793814, + 0.0043808212503790855, + 0.003222229890525341, + 0.0020218840800225735, + 0.00660955673083663, + 0.005384801421314478, + 0.009086893871426582, + 0.006490906700491905, + 0.008402764797210693, + 0.008849861100316048, + 0.008753003552556038, + 0.005676451604813337, + 0.005007306579500437, + 0.006467747502028942, + 0.006428238935768604, + 0.008318303152918816, + 0.0074117169715464115, + 0.006087689660489559, + 0.003537186188623309, + 0.0039038127288222313, + 0.004497076850384474, + 0.005290928762406111, + 0.003576885210350156, + 0.0035673994570970535, + 0.005124933551996946, + 0.006333056837320328, + 0.003334259847179055, + 0.0031110248528420925, + 0.003199595957994461, + 0.002879708306863904, + 0.0026108145248144865, + 0.0027519601862877607, + 0.0030098827555775642, + 0.003939747344702482, + 0.0022631140891462564, + 0.0027048331685364246, + 0.24909880757331848, + 0.08440440893173218, + 0.21781989932060242, + 0.02619631588459015, + 0.0133185600861907, + 0.015947580337524414, + 0.007410677615553141, + 0.01636327989399433, + 0.004973683040589094, + 0.008924048393964767, + 0.005318699404597282, + 0.004507073666900396, + 0.005683936644345522, + 0.004599571228027344, + 0.003627046477049589, + 0.005186470225453377, + 0.0035544501151889563, + 0.003990064840763807, + 0.007001896854490042, + 0.009347766637802124, + 0.006326339207589626, + 0.0037031120155006647, + 0.002884184243157506, + 0.0053916252218186855, + 0.004664169624447823, + 0.003762387204915285, + 0.0029461621306836605, + 0.0027212854474782944, + 0.0033752804156392813, + 0.0033108338247984648, + 0.002505149459466338, + 0.003215289209038019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.009780817665159702 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002488613361492753 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.002488613361492753, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002488613361492753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14913180470466614, + 0.04991405829787254, + 0.06088607385754585, + 0.0302300862967968, + 0.05623333528637886, + 0.03427053242921829, + 0.16570186614990234, + 0.044676560908555984, + 0.36183372139930725, + 0.04712199047207832 + ], + "credit_map": [ + 0.018802998587489128, + 0.006825149524956942, + 0.05631231144070625, + 0.01342347078025341, + 0.008193068206310272, + 0.010716579854488373, + 0.009909199550747871, + 0.024949023500084877, + 0.005442784167826176, + 0.003713183104991913, + 0.01116862054914236, + 0.006284418515861034, + 0.004966911859810352, + 0.00820937566459179, + 0.006477673072367907, + 0.0036510915961116552, + 0.0074464199133217335, + 0.004892169497907162, + 0.014907716773450375, + 0.007043671794235706, + 0.006368317641317844, + 0.0070687527768313885, + 0.006274053826928139, + 0.006884975358843803, + 0.0031767133623361588, + 0.004181774333119392, + 0.004511035513132811, + 0.005439502652734518, + 0.004008651245385408, + 0.003902998985722661, + 0.0024645444937050343, + 0.00254486664198339, + 0.013919001445174217, + 0.01170773059129715, + 0.005118764471262693, + 0.004950474016368389, + 0.004314159508794546, + 0.007549608126282692, + 0.004817357752472162, + 0.00385624123737216, + 0.004649046342819929, + 0.005957275629043579, + 0.0031501760240644217, + 0.002665493404492736, + 0.0055797891691327095, + 0.004850354976952076, + 0.0032037366181612015, + 0.004214661195874214, + 0.06406879425048828, + 0.056988425552845, + 0.01563766784965992, + 0.006762153003364801, + 0.00767373526468873, + 0.007928884588181973, + 0.003123317379504442, + 0.0035188794136047363, + 0.0038432555738836527, + 0.00877869687974453, + 0.004801296163350344, + 0.005336334463208914, + 0.008375531993806362, + 0.005255374126136303, + 0.003904524026438594, + 0.004381549544632435, + 0.024736454710364342, + 0.03859412670135498, + 0.07071969658136368, + 0.09705657511949539, + 0.03973934054374695, + 0.025674762204289436, + 0.018244469538331032, + 0.04706830158829689, + 0.006157370284199715, + 0.005956229753792286, + 0.003934805281460285, + 0.004252826329320669, + 0.010008793324232101, + 0.0062138596549630165, + 0.00605821842327714, + 0.004539885558187962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.002488613361492753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, high-angle, overhead shot slowly pulls back to reveal a magnificent spectacle: a boundless sea of golden clouds roiling violently, like a seething ocean of molten gold. Vast waves collide, forming deep eddies and towering crests. At the center of this violent and sacred scene, the Monkey King, clad in a magnificent suit of armor, stands proud. He wears a phoenix-winged, purple-gold crown, its feathers trembling in the strong wind. He clad in golden chain mail, each plate reflecting the almighty golden light from the heavens, shimmering with brilliance. His feet, clad in lotus-silk, cloud-walking shoes, firmly anchor him above the turbulent sea of clouds. His powerful arm grips the heavy Ruyi Jingu Staff, one end of which is deeply embedded in the swirling clouds. Like a stabilizing force, his figure remains unshaken, regardless of the turbulent waves around him. His signature red cape dances wildly behind him, creating a striking contrast against the golden backdrop. Although it is viewed from a high altitude, one can still feel his unruly aura and his eyes as sharp as torches, as if they have penetrated through layers of clouds and mist, gazing into the unknown distance. The entire picture is filled with a mythological atmosphere of grandeur, solemnity and endless power.", + "seed": 7478, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_1/index_8105_seed_7478_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04499611537903547, + "vlm_logprob": -0.04499611537903547, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.04499611537903547, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19146139919757843 + ], + "opposite_logprob_tokens": [ + -1.816461443901062 + ], + "logprob_reward": -0.19146139919757843, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19146139919757843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18510045111179352, + 0.053802069276571274, + 0.051523394882678986, + 0.06495129317045212, + 0.27432340383529663, + 0.11165203899145126, + 0.044164080172777176, + 0.10563858598470688, + 0.04660589247941971, + 0.06223874166607857 + ], + "credit_map": [ + 0.018601177260279655, + 0.022528018802404404, + 0.03364923968911171, + 0.014000499621033669, + 0.023627204820513725, + 0.02785792201757431, + 0.017835913226008415, + 0.027000470086932182, + 0.005320692900568247, + 0.012776246294379234, + 0.005900430027395487, + 0.006668800488114357, + 0.007727410178631544, + 0.005038294475525618, + 0.0038834153674542904, + 0.0064867837354540825, + 0.0060246167704463005, + 0.007736853789538145, + 0.0075744218192994595, + 0.0049765692092478275, + 0.006688087712973356, + 0.008221795782446861, + 0.0037093041464686394, + 0.006591747514903545, + 0.010090063326060772, + 0.00633395416662097, + 0.01103993784636259, + 0.005596264731138945, + 0.006520096678286791, + 0.011732316575944424, + 0.004103781655430794, + 0.009534882381558418, + 0.04837926849722862, + 0.027341315522789955, + 0.041786063462495804, + 0.02075447328388691, + 0.02083497866988182, + 0.043032534420490265, + 0.02182614430785179, + 0.05036862567067146, + 0.007820846512913704, + 0.006129983812570572, + 0.01369174662977457, + 0.007233998738229275, + 0.008479633368551731, + 0.03608446940779686, + 0.00771702267229557, + 0.024494342505931854, + 0.006525604985654354, + 0.003155299462378025, + 0.005164253991097212, + 0.0034669600427150726, + 0.002875276142731309, + 0.009967857040464878, + 0.0025517167523503304, + 0.010457109659910202, + 0.012136085890233517, + 0.010884345509111881, + 0.01723751239478588, + 0.010500394739210606, + 0.012392900884151459, + 0.01697191223502159, + 0.014296078123152256, + 0.011219357140362263, + 0.004090947099030018, + 0.003133094636723399, + 0.0053796181455254555, + 0.005565870553255081, + 0.007473072037100792, + 0.007620895281434059, + 0.00886041671037674, + 0.004481982905417681, + 0.006251942832022905, + 0.003848502179607749, + 0.0039028951432555914, + 0.003860080847516656, + 0.006255764979869127, + 0.016965297982096672, + 0.0062184762209653854, + 0.014935780316591263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.19146139919757843 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005272056441754103 + ], + "opposite_logprob_tokens": [ + -5.255271911621094 + ], + "logprob_reward": -0.005272056441754103, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005272056441754103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48117494583129883, + 0.08634155243635178, + 0.0660533756017685, + 0.055272914469242096, + 0.12969976663589478, + 0.04662538319826126, + 0.030066777020692825, + 0.04736989736557007, + 0.033362336456775665, + 0.024033093824982643 + ], + "credit_map": [ + 0.05109037831425667, + 0.03923754766583443, + 0.10631836950778961, + 0.044045720249414444, + 0.06291902810335159, + 0.05870090797543526, + 0.046661797910928726, + 0.07220117747783661, + 0.010255303233861923, + 0.010839554481208324, + 0.019489655271172523, + 0.007303405087441206, + 0.005626127123832703, + 0.009431922808289528, + 0.004467346705496311, + 0.018928224220871925, + 0.008073324337601662, + 0.009961143136024475, + 0.012998024001717567, + 0.005705499090254307, + 0.006380925420671701, + 0.009375630877912045, + 0.0036252897698432207, + 0.009933529421687126, + 0.008480566553771496, + 0.004650263115763664, + 0.0077666025608778, + 0.004823397845029831, + 0.005599919706583023, + 0.012338799424469471, + 0.003952078055590391, + 0.007661282550543547, + 0.02376202866435051, + 0.013435963541269302, + 0.01728096231818199, + 0.00984616857022047, + 0.007597774732857943, + 0.01996000111103058, + 0.009563632309436798, + 0.028253212571144104, + 0.0034724248107522726, + 0.0045062773860991, + 0.004758798982948065, + 0.003979082684963942, + 0.0038459734059870243, + 0.013722838833928108, + 0.004579068161547184, + 0.007760913111269474, + 0.003595343790948391, + 0.0046699801459908485, + 0.0033350714948028326, + 0.001820540288463235, + 0.0027551965322345495, + 0.006138151977211237, + 0.002317778067663312, + 0.005434710066765547, + 0.006016812287271023, + 0.006741099525243044, + 0.006460016593337059, + 0.0047486936673521996, + 0.005476200487464666, + 0.00618773465976119, + 0.006742728874087334, + 0.004996604286134243, + 0.003740407759323716, + 0.002831553341820836, + 0.004114815965294838, + 0.0037339047994464636, + 0.004810789600014687, + 0.0053374432027339935, + 0.005082561634480953, + 0.003710854332894087, + 0.001601592288352549, + 0.0007801794563420117, + 0.0029666845221072435, + 0.001272027031518519, + 0.0018589464016258717, + 0.008106173016130924, + 0.00193825236056, + 0.005509234964847565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.005272056441754103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193184673786163 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193184673786163, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193184673786163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29160696268081665, + 0.11993656307458878, + 0.20159697532653809, + 0.07197269052267075, + 0.05258363112807274, + 0.0503033809363842, + 0.07408551871776581, + 0.018846355378627777, + 0.06846728175878525, + 0.05060058459639549 + ], + "credit_map": [ + 0.027248293161392212, + 0.026515016332268715, + 0.06327370554208755, + 0.02604677341878414, + 0.035242266952991486, + 0.03032960183918476, + 0.0192621648311615, + 0.06368916481733322, + 0.01048986054956913, + 0.011566577479243279, + 0.02771749719977379, + 0.011227431707084179, + 0.010837267152965069, + 0.011289318092167377, + 0.010616938583552837, + 0.026191676035523415, + 0.020379500463604927, + 0.04017896577715874, + 0.022982686758041382, + 0.02256486751139164, + 0.0322142168879509, + 0.02294931560754776, + 0.017037097364664078, + 0.023290352895855904, + 0.010440285317599773, + 0.007801821455359459, + 0.009027400054037571, + 0.006236338522285223, + 0.00627256790176034, + 0.0162223968654871, + 0.005430230870842934, + 0.01054165419191122, + 0.008377059362828732, + 0.006077114026993513, + 0.007483610417693853, + 0.005329072941094637, + 0.004216624889522791, + 0.007219781633466482, + 0.004573660437017679, + 0.009306714870035648, + 0.005005907732993364, + 0.004184338264167309, + 0.006682443432509899, + 0.00484973331913352, + 0.004435784183442593, + 0.01148888561874628, + 0.005380313843488693, + 0.00827598012983799, + 0.008942992426455021, + 0.008283211849629879, + 0.007556240074336529, + 0.0055411141365766525, + 0.005165969021618366, + 0.017038600519299507, + 0.006787305697798729, + 0.014770090579986572, + 0.002334177726879716, + 0.001991985132917762, + 0.0035143319983035326, + 0.0018527654465287924, + 0.0018375263316556811, + 0.0024049864150583744, + 0.002229254227131605, + 0.002681329380720854, + 0.006824469659477472, + 0.005429488606750965, + 0.008144129998981953, + 0.00793963298201561, + 0.01057420764118433, + 0.011105573736131191, + 0.01137496903538704, + 0.007074810564517975, + 0.004001640249043703, + 0.002237023552879691, + 0.0041755447164177895, + 0.0028290394693613052, + 0.003946393262594938, + 0.015880821272730827, + 0.004297650884836912, + 0.013232474215328693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014193184673786163 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541004456579685 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012541004456579685, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541004456579685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20494344830513, + 0.10001470893621445, + 0.06331346184015274, + 0.06823885440826416, + 0.3674549162387848, + 0.028777875006198883, + 0.016076313331723213, + 0.02788473479449749, + 0.08565936982631683, + 0.03763631358742714 + ], + "credit_map": [ + 0.020294368267059326, + 0.011128506623208523, + 0.06524438410997391, + 0.016074076294898987, + 0.010590590536594391, + 0.01767934486269951, + 0.011662796139717102, + 0.052269384264945984, + 0.006916332989931107, + 0.01591319963335991, + 0.015359941869974136, + 0.006763411685824394, + 0.007915244437754154, + 0.008987165987491608, + 0.007489755749702454, + 0.030669664964079857, + 0.008058395236730576, + 0.012137374840676785, + 0.00812558177858591, + 0.005915068089962006, + 0.007100158836692572, + 0.007599313743412495, + 0.004147058352828026, + 0.010230515152215958, + 0.00820887554436922, + 0.013034590519964695, + 0.009165390394628048, + 0.006291608791798353, + 0.008250867016613483, + 0.01179295964539051, + 0.004880091641098261, + 0.006614469923079014, + 0.05248657613992691, + 0.056115731596946716, + 0.05441351234912872, + 0.045630406588315964, + 0.034240517765283585, + 0.04463588073849678, + 0.03264976292848587, + 0.047282520681619644, + 0.003244539024308324, + 0.0027766539715230465, + 0.003220900194719434, + 0.00328079960308969, + 0.00496271625161171, + 0.0036575833801180124, + 0.0053288317285478115, + 0.002305852947756648, + 0.001928370096720755, + 0.0016899330075830221, + 0.0028155772015452385, + 0.001724562724120915, + 0.0019572903402149677, + 0.0021957054268568754, + 0.001549681299366057, + 0.002215194283053279, + 0.0033445535227656364, + 0.003101933980360627, + 0.004498120862990618, + 0.003083295887336135, + 0.003262340323999524, + 0.0038800223264843225, + 0.003507566172629595, + 0.0032069047447293997, + 0.00857405923306942, + 0.005698591936379671, + 0.010126358829438686, + 0.010025311261415482, + 0.013762539252638817, + 0.01492470595985651, + 0.013867772184312344, + 0.008680031634867191, + 0.002213270403444767, + 0.0009986102813854814, + 0.003696694504469633, + 0.0015570098767057061, + 0.0013955597532913089, + 0.0147054148837924, + 0.0020815597381442785, + 0.010988199152052402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.012541004456579685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a high-angle overhead view that slowly pulls back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a boundless sea of golden clouds with violent waves, eddies, and towering crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Monkey King visible wearing a phoenix-winged purple-gold crown and golden chain mail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the Monkey King holding a Ruyi Jingu Staff with one end embedded in the swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the Monkey King's red cape visible dancing wildly behind him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001512932125478983 + ], + "opposite_logprob_tokens": [ + -6.5015130043029785 + ], + "logprob_reward": -0.001512932125478983, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001512932125478983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23881219327449799, + 0.07658383995294571, + 0.07955269515514374, + 0.07197225093841553, + 0.10298965871334076, + 0.08435115218162537, + 0.08684872090816498, + 0.05480840429663658, + 0.1092490553855896, + 0.09483202546834946 + ], + "credit_map": [ + 0.02500048466026783, + 0.016635030508041382, + 0.06459915637969971, + 0.02051282301545143, + 0.01802029088139534, + 0.01977938413619995, + 0.020448898896574974, + 0.05381613224744797, + 0.0067026326432824135, + 0.013461073860526085, + 0.009926063008606434, + 0.008228878490626812, + 0.009193236008286476, + 0.007984316907823086, + 0.0073566678911447525, + 0.013730972073972225, + 0.00901778507977724, + 0.011511920019984245, + 0.011647161096334457, + 0.006614618934690952, + 0.009761233814060688, + 0.010931380093097687, + 0.005408499389886856, + 0.014660097658634186, + 0.007120365742594004, + 0.009797170758247375, + 0.012233379296958447, + 0.005585664417594671, + 0.00925256498157978, + 0.014929489232599735, + 0.006032831966876984, + 0.007020790595561266, + 0.01821395941078663, + 0.012564290314912796, + 0.01761534810066223, + 0.008363863453269005, + 0.0067021106369793415, + 0.013703756965696812, + 0.00791226327419281, + 0.0179140567779541, + 0.007445202209055424, + 0.006072843447327614, + 0.01239614374935627, + 0.006292921956628561, + 0.005755970254540443, + 0.02451268769800663, + 0.007312231231480837, + 0.014563150703907013, + 0.008752669207751751, + 0.007179931737482548, + 0.01282504666596651, + 0.00489790178835392, + 0.004926488269120455, + 0.026839084923267365, + 0.006189436186105013, + 0.015238163992762566, + 0.006967886351048946, + 0.005181814543902874, + 0.011713068000972271, + 0.004539028275758028, + 0.0045914482325315475, + 0.01025809533894062, + 0.005836202763020992, + 0.005720856599509716, + 0.011898416094481945, + 0.007434528321027756, + 0.011544258333742619, + 0.01320705097168684, + 0.01552137453109026, + 0.020131753757596016, + 0.01576687954366207, + 0.013744798488914967, + 0.010460954159498215, + 0.0089041693136096, + 0.009488188661634922, + 0.008199234493076801, + 0.01087130419909954, + 0.01753963530063629, + 0.009942394681274891, + 0.019426148384809494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.001512932125478983 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_10/index_tensor([6322, 6322])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_10/index_tensor([6322, 6322])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b42a53899cc21e782d803a1b4b5f494973315894 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_10/index_tensor([6322, 6322])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, taken from a slight overhead angle, focuses on a chubby orange tabby cat nestled in an interior space with polished light-brown wood floors. Warm afternoon sunlight slants in from outside the frame, casting a bright patch of light on the floor. Within this patch of light, the cat, arching its chubby body, begins a series of rapid, clumsy spins, pivoting on its front legs. Concentrated in pursuit of its long, furry tail, its movements appear uncoordinated due to its sleek shape, each spin imbued with a comical wobbling quality. Its short, sturdy limbs struggle to maintain balance, yet its claws repeatedly reach out in vain. The sunlight illuminates its fluffy, soft fur, casting a golden halo around the edges. Its green eyes, fixed intently on the tip of its tail, which it can't grasp, exude a sense of both persistence and innocence. The warm tones of the entire scene create a deeply warm, relaxed, and playful atmosphere.", + "seed": 3043, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_10/index_6322_seed_3043_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16118726679123938, + "vlm_logprob": -0.16118726679123938, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.16118726679123938, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010156379081308842 + ], + "opposite_logprob_tokens": [ + -4.760156154632568 + ], + "logprob_reward": -0.010156379081308842, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010156379081308842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08096325397491455, + 0.11235310137271881, + 0.2114579826593399, + 0.1029454916715622, + 0.14091706275939941, + 0.08867684751749039, + 0.07263103127479553, + 0.11730311810970306, + 0.0375978983938694, + 0.0351543053984642 + ], + "credit_map": [ + 0.009649028070271015, + 0.01024733018130064, + 0.009278429672122002, + 0.012763760052621365, + 0.009432483464479446, + 0.005507367197424173, + 0.011434067972004414, + 0.01265077106654644, + 0.012531843967735767, + 0.01787158101797104, + 0.013623438775539398, + 0.010999101214110851, + 0.01669723354279995, + 0.008108887821435928, + 0.016969852149486542, + 0.015551147982478142, + 0.04704723134636879, + 0.008857914246618748, + 0.07469606399536133, + 0.010664763860404491, + 0.012237897142767906, + 0.009381859563291073, + 0.01623164862394333, + 0.03234058991074562, + 0.012637556530535221, + 0.008579244837164879, + 0.012229712679982185, + 0.01176151167601347, + 0.013438182882964611, + 0.010305248200893402, + 0.014840028248727322, + 0.01915399543941021, + 0.032862547785043716, + 0.011797730810940266, + 0.029550619423389435, + 0.010910994373261929, + 0.021077880635857582, + 0.007562737911939621, + 0.0110299838706851, + 0.016124535351991653, + 0.00618047034367919, + 0.008565204218029976, + 0.015133798122406006, + 0.008761269971728325, + 0.010687497444450855, + 0.01032460667192936, + 0.013089652173221111, + 0.015934336930513382, + 0.00563160702586174, + 0.007821244187653065, + 0.006334010045975447, + 0.007510501425713301, + 0.017179282382130623, + 0.005973842926323414, + 0.015067231841385365, + 0.007113303057849407, + 0.008496928960084915, + 0.011672786436975002, + 0.008098622784018517, + 0.012291213497519493, + 0.01719210110604763, + 0.029291711747646332, + 0.015107361599802971, + 0.015152371488511562, + 0.0038640869315713644, + 0.005633911117911339, + 0.005204289685934782, + 0.004075275268405676, + 0.0046197958290576935, + 0.004288614261895418, + 0.005995477084070444, + 0.003916445188224316, + 0.004206858109682798, + 0.0036422614939510822, + 0.006289481185376644, + 0.00455139484256506, + 0.0036016758531332016, + 0.0018826921004801989, + 0.0068862405605614185, + 0.004093698225915432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.010156379081308842 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031986285466700792 + ], + "opposite_logprob_tokens": [ + -5.753198623657227 + ], + "logprob_reward": -0.0031986285466700792, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031986285466700792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10033571720123291, + 0.07335292547941208, + 0.07757287472486496, + 0.10868223756551743, + 0.08349351584911346, + 0.23381419479846954, + 0.13950397074222565, + 0.07701223343610764, + 0.051646023988723755, + 0.05458629131317139 + ], + "credit_map": [ + 0.012454629875719547, + 0.012650465592741966, + 0.0158463716506958, + 0.01544736698269844, + 0.010668113827705383, + 0.007965539582073689, + 0.010248981416225433, + 0.015054243616759777, + 0.010018558241426945, + 0.007568527478724718, + 0.007994260638952255, + 0.007975972257554531, + 0.010350828059017658, + 0.009495683014392853, + 0.008210621774196625, + 0.01173846609890461, + 0.009044982492923737, + 0.006724715232849121, + 0.01318912860006094, + 0.006429960485547781, + 0.00718692597001791, + 0.0073149483650922775, + 0.010871033184230328, + 0.016811178997159004, + 0.015248375944793224, + 0.008272947743535042, + 0.01702362671494484, + 0.012288639321923256, + 0.015536757186055183, + 0.009299496188759804, + 0.015007607638835907, + 0.0160047747194767, + 0.010729413479566574, + 0.005291877314448357, + 0.012338928878307343, + 0.005416186060756445, + 0.011384015902876854, + 0.009255148470401764, + 0.008692997507750988, + 0.02038494311273098, + 0.014675694517791271, + 0.023302799090743065, + 0.06146518513560295, + 0.019023125991225243, + 0.030216719955205917, + 0.02061031572520733, + 0.026877185329794884, + 0.03764314949512482, + 0.010915027931332588, + 0.013696335256099701, + 0.025093544274568558, + 0.014455706812441349, + 0.0191934984177351, + 0.013580767437815666, + 0.02711891010403633, + 0.015450173057615757, + 0.005925771780312061, + 0.0064957658760249615, + 0.010173894464969635, + 0.006187519058585167, + 0.008389945141971111, + 0.021419532597064972, + 0.009096677415072918, + 0.00932311825454235, + 0.004336211830377579, + 0.005913183093070984, + 0.009070094674825668, + 0.004467438440769911, + 0.009458901360630989, + 0.006404641550034285, + 0.006285923067480326, + 0.005709625780582428, + 0.0065107932314276695, + 0.005044056102633476, + 0.009738290682435036, + 0.006989792920649052, + 0.007051875349134207, + 0.003303894307464361, + 0.00955155212432146, + 0.00639603566378355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.0031986285466700792 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578182220458984 + ], + "opposite_logprob_tokens": [ + -0.6328182220458984 + ], + "logprob_reward": -0.7578182220458984, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578182220458984, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06860513985157013, + 0.03078520856797695, + 0.1301429122686386, + 0.10746455937623978, + 0.10676611214876175, + 0.1738317906856537, + 0.19537104666233063, + 0.030465923249721527, + 0.0905274823307991, + 0.06603983789682388 + ], + "credit_map": [ + 0.008192487061023712, + 0.008555835112929344, + 0.007938297465443611, + 0.011743729934096336, + 0.008227312937378883, + 0.004423422738909721, + 0.009698763489723206, + 0.00982529204338789, + 0.003274019807577133, + 0.002790022874251008, + 0.004457541275769472, + 0.0032034297473728657, + 0.004322977736592293, + 0.0037940179463475943, + 0.004450363107025623, + 0.0044928365387022495, + 0.021213294938206673, + 0.007022283039987087, + 0.038383737206459045, + 0.009177627973258495, + 0.009707427583634853, + 0.00846018921583891, + 0.014699526131153107, + 0.02147882804274559, + 0.014646287076175213, + 0.008422051556408405, + 0.013981631025671959, + 0.012003849260509014, + 0.01618853583931923, + 0.008405245840549469, + 0.016259491443634033, + 0.01755746826529503, + 0.012291102670133114, + 0.009246472269296646, + 0.014122010208666325, + 0.008289911784231663, + 0.01537729799747467, + 0.01117768231779337, + 0.009912730194628239, + 0.026348909363150597, + 0.020138965919613838, + 0.017419476062059402, + 0.017040902748703957, + 0.015981703996658325, + 0.024309596046805382, + 0.027765586972236633, + 0.029523126780986786, + 0.02165244333446026, + 0.015165790915489197, + 0.023767447099089622, + 0.02165224961936474, + 0.021926578134298325, + 0.04439470171928406, + 0.014645458199083805, + 0.03188781812787056, + 0.02193102054297924, + 0.0032317042350769043, + 0.002908645896241069, + 0.0037307117599993944, + 0.0033879403490573168, + 0.003805693006142974, + 0.0069260806776583195, + 0.0033275559544563293, + 0.0031475904397666454, + 0.007621744181960821, + 0.008375885896384716, + 0.01534227654337883, + 0.009101592935621738, + 0.015917493030428886, + 0.012442462146282196, + 0.011669954285025597, + 0.01005607470870018, + 0.006566238589584827, + 0.0070600793696939945, + 0.007097745314240456, + 0.0077036526054143906, + 0.008895021863281727, + 0.009467550553381443, + 0.009141949936747551, + 0.010107601061463356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.7578182220458984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191773720085621 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.014191773720085621, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191773720085621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12924276292324066, + 0.10773243755102158, + 0.29165706038475037, + 0.10431727021932602, + 0.05220161750912666, + 0.08838801085948944, + 0.08289802819490433, + 0.05624513700604439, + 0.03555124253034592, + 0.05176642909646034 + ], + "credit_map": [ + 0.01706671342253685, + 0.013637627474963665, + 0.016781240701675415, + 0.01680718921124935, + 0.017724044620990753, + 0.010268038138747215, + 0.01724805124104023, + 0.019709855318069458, + 0.015220142900943756, + 0.012707313522696495, + 0.017717337235808372, + 0.009935863316059113, + 0.01270544808357954, + 0.014704480767250061, + 0.008548605255782604, + 0.016193248331546783, + 0.06930309534072876, + 0.015334551222622395, + 0.10098867118358612, + 0.01923413574695587, + 0.017211614176630974, + 0.014999809674918652, + 0.0207113865762949, + 0.033873818814754486, + 0.011789320968091488, + 0.010083573870360851, + 0.01210861001163721, + 0.015885217115283012, + 0.014182611368596554, + 0.01202176883816719, + 0.011263218708336353, + 0.01698295585811138, + 0.005369497463107109, + 0.005729327443987131, + 0.008104232139885426, + 0.004474538378417492, + 0.008202060125768185, + 0.004727046471089125, + 0.006664210464805365, + 0.008930706419050694, + 0.009674480184912682, + 0.008939116261899471, + 0.010648978874087334, + 0.007111500017344952, + 0.014684253372251987, + 0.013667422346770763, + 0.013470870442688465, + 0.010191386565566063, + 0.006066122092306614, + 0.010893910191953182, + 0.006056603044271469, + 0.009234446100890636, + 0.02024301514029503, + 0.006946571171283722, + 0.014375412836670876, + 0.009081941097974777, + 0.004627905320376158, + 0.005345901008695364, + 0.006300546228885651, + 0.006750774569809437, + 0.006337805651128292, + 0.012873721309006214, + 0.007201798725873232, + 0.006806683260947466, + 0.0031388753559440374, + 0.004613248631358147, + 0.0034501859918236732, + 0.0037541391793638468, + 0.005678596906363964, + 0.004763839766383171, + 0.00527908094227314, + 0.004873274359852076, + 0.005746672861278057, + 0.006995389703661203, + 0.004857007879763842, + 0.0066842311061918736, + 0.007661753799766302, + 0.004748630337417126, + 0.007509348914027214, + 0.00756339356303215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.014191773720085621 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571330562233925 + ], + "opposite_logprob_tokens": [ + -3.895571231842041 + ], + "logprob_reward": -0.020571330562233925, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571330562233925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10260086506605148, + 0.1010633260011673, + 0.09257645905017853, + 0.0939602479338646, + 0.051640935242176056, + 0.11936519294977188, + 0.10509122908115387, + 0.1649821251630783, + 0.03624594956636429, + 0.13247370719909668 + ], + "credit_map": [ + 0.01260625384747982, + 0.012682268396019936, + 0.01377277821302414, + 0.016342436894774437, + 0.011468455195426941, + 0.007952812127768993, + 0.012790056876838207, + 0.014985805377364159, + 0.01402345858514309, + 0.008784458041191101, + 0.022090978920459747, + 0.014377444051206112, + 0.010367944836616516, + 0.006298141088336706, + 0.011143715120851994, + 0.013977179303765297, + 0.015189032070338726, + 0.005501009523868561, + 0.028582751750946045, + 0.005929556675255299, + 0.006459008902311325, + 0.00646783085539937, + 0.009116566739976406, + 0.015330699272453785, + 0.012844674289226532, + 0.007664409466087818, + 0.013959026895463467, + 0.010563013143837452, + 0.01113707572221756, + 0.009875933639705181, + 0.011292915791273117, + 0.01662319526076317, + 0.011311209760606289, + 0.002883469918742776, + 0.00561022013425827, + 0.004052815493196249, + 0.006029749289155006, + 0.005377624183893204, + 0.003434115555137396, + 0.01294173114001751, + 0.008295021951198578, + 0.012601473368704319, + 0.013285179622471333, + 0.010589388199150562, + 0.02333679050207138, + 0.015339819714426994, + 0.018700066953897476, + 0.017217449843883514, + 0.008106105960905552, + 0.012471719644963741, + 0.011685146950185299, + 0.012034392915666103, + 0.02176971174776554, + 0.011769702658057213, + 0.015492161735892296, + 0.011762292124330997, + 0.020580673590302467, + 0.01668843999505043, + 0.02257283218204975, + 0.015193351544439793, + 0.017617173492908478, + 0.03560538962483406, + 0.018985340371727943, + 0.01773892715573311, + 0.0034495238214731216, + 0.0032565973233431578, + 0.0037767039611935616, + 0.0037734336219727993, + 0.007118146400898695, + 0.00409501139074564, + 0.006737167946994305, + 0.004039364866912365, + 0.01665116846561432, + 0.012127849273383617, + 0.02175810933113098, + 0.016757585108280182, + 0.01626504398882389, + 0.010285579599440098, + 0.021229615435004234, + 0.01739875227212906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.020571330562233925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, taken from a slight overhead angle, focuses on a chubby orange tabby cat nestled in an interior space with polished light-brown wood floors. Warm afternoon sunlight slants in from outside the frame, casting a bright patch of light on the floor. Within this patch of light, the cat, arching its chubby body, begins a series of rapid, clumsy spins, pivoting on its front legs. Concentrated in pursuit of its long, furry tail, its movements appear uncoordinated due to its sleek shape, each spin imbued with a comical wobbling quality. Its short, sturdy limbs struggle to maintain balance, yet its claws repeatedly reach out in vain. The sunlight illuminates its fluffy, soft fur, casting a golden halo around the edges. Its green eyes, fixed intently on the tip of its tail, which it can't grasp, exude a sense of both persistence and innocence. The warm tones of the entire scene create a deeply warm, relaxed, and playful atmosphere.", + "seed": 3044, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_10/index_6322_seed_3044_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024806815036572515, + "vlm_logprob": -0.024806815036572515, + "vlm_yes_no_margin": 4.249999952316284, + "vlm_token_credit": -0.024806815036572515, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015723537653684616 + ], + "opposite_logprob_tokens": [ + -4.390723705291748 + ], + "logprob_reward": -0.015723537653684616, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015723537653684616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15229745209217072, + 0.05710083991289139, + 0.0593855045735836, + 0.11425912380218506, + 0.0981285572052002, + 0.08496133983135223, + 0.15276657044887543, + 0.02848242223262787, + 0.19756639003753662, + 0.05505182966589928 + ], + "credit_map": [ + 0.015596131794154644, + 0.019808953627943993, + 0.01858615316450596, + 0.02895207144320011, + 0.013422863557934761, + 0.02350441738963127, + 0.019462890923023224, + 0.012963972054421902, + 0.0057052550837397575, + 0.0033800085075199604, + 0.008716762065887451, + 0.005150104872882366, + 0.005825890228152275, + 0.01600772514939308, + 0.005338536109775305, + 0.006976558361202478, + 0.005106375087052584, + 0.007013048976659775, + 0.009121857583522797, + 0.00550464540719986, + 0.006897845305502415, + 0.015354802832007408, + 0.0044356281869113445, + 0.005951299332082272, + 0.02279485948383808, + 0.008683163672685623, + 0.011841748841106892, + 0.01214517466723919, + 0.008040652610361576, + 0.031167559325695038, + 0.009696927852928638, + 0.00988903921097517, + 0.01032937504351139, + 0.00941707007586956, + 0.012612005695700645, + 0.010240080766379833, + 0.007231525611132383, + 0.01316581666469574, + 0.011047059670090675, + 0.024085627868771553, + 0.00830337218940258, + 0.006631362251937389, + 0.012626220472157001, + 0.007264749146997929, + 0.01216532289981842, + 0.012129795737564564, + 0.01720452681183815, + 0.008635996840894222, + 0.021574417129158974, + 0.01360122486948967, + 0.01992235705256462, + 0.01855420507490635, + 0.015182146802544594, + 0.036462511867284775, + 0.015008019283413887, + 0.012461679056286812, + 0.0047809951938688755, + 0.003674836363643408, + 0.005893269088119268, + 0.0030839224345982075, + 0.002752471249550581, + 0.0035390956327319145, + 0.0025535349268466234, + 0.0022042992059141397, + 0.024615133181214333, + 0.02950831688940525, + 0.03238696604967117, + 0.014410292729735374, + 0.030507966876029968, + 0.025959236547350883, + 0.02403298206627369, + 0.01614551991224289, + 0.008180608041584492, + 0.010632309131324291, + 0.01018880121409893, + 0.008722307160496712, + 0.003620374482125044, + 0.00427037151530385, + 0.0048748161643743515, + 0.004562241956591606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.015723537653684616 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001948602613992989 + ], + "opposite_logprob_tokens": [ + -6.251948833465576 + ], + "logprob_reward": -0.001948602613992989, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001948602613992989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13605369627475739, + 0.05941970646381378, + 0.07732301950454712, + 0.08827144652605057, + 0.24623258411884308, + 0.0451408252120018, + 0.062458403408527374, + 0.042495518922805786, + 0.08841373771429062, + 0.15419107675552368 + ], + "credit_map": [ + 0.013386816717684269, + 0.009039542637765408, + 0.016173668205738068, + 0.01632070541381836, + 0.009872226044535637, + 0.018493665382266045, + 0.039924293756484985, + 0.01284277718514204, + 0.005760623142123222, + 0.009478211402893066, + 0.009048527106642723, + 0.0064292545430362225, + 0.006448198575526476, + 0.01032568421214819, + 0.0055906884372234344, + 0.006338512524962425, + 0.009884997271001339, + 0.010317759588360786, + 0.012701699510216713, + 0.009102925658226013, + 0.007839641533792019, + 0.01101462822407484, + 0.007561427541077137, + 0.008899930864572525, + 0.009141551330685616, + 0.007095797918736935, + 0.020882820710539818, + 0.007556608412414789, + 0.005056299734860659, + 0.02575620822608471, + 0.00512701366096735, + 0.007655137684196234, + 0.025712035596370697, + 0.019824955612421036, + 0.03365134075284004, + 0.020961055532097816, + 0.015514006838202477, + 0.040780775249004364, + 0.02535327523946762, + 0.06443510949611664, + 0.00598238967359066, + 0.002687026746571064, + 0.008052901364862919, + 0.003998694475740194, + 0.004647036548703909, + 0.009255681186914444, + 0.004573180805891752, + 0.005943909753113985, + 0.008341226726770401, + 0.005228579510003328, + 0.007367829792201519, + 0.006707764230668545, + 0.006333933211863041, + 0.013873504474759102, + 0.006812042556703091, + 0.007793516386300325, + 0.006302943918853998, + 0.00387569866143167, + 0.008731992915272713, + 0.007346800994127989, + 0.0029460431542247534, + 0.0046557229943573475, + 0.004968793131411076, + 0.003667516866698861, + 0.008893835358321667, + 0.015541739761829376, + 0.011552843265235424, + 0.009389539249241352, + 0.01368672214448452, + 0.010344934649765491, + 0.01029701717197895, + 0.008707093074917793, + 0.023038316518068314, + 0.0271709393709898, + 0.028089327737689018, + 0.026754263788461685, + 0.012285170145332813, + 0.010835289023816586, + 0.014848317950963974, + 0.01116944383829832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.001948602613992989 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001003623008728 + ], + "opposite_logprob_tokens": [ + -2.69500994682312 + ], + "logprob_reward": -0.07001003623008728, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001003623008728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09823934733867645, + 0.04432791471481323, + 0.07300347834825516, + 0.04414437338709831, + 0.04653220251202583, + 0.1324651837348938, + 0.05519554391503334, + 0.14348401129245758, + 0.29705819487571716, + 0.06554977595806122 + ], + "credit_map": [ + 0.00920474249869585, + 0.011959939263761044, + 0.01156378909945488, + 0.011599928140640259, + 0.013020738959312439, + 0.01082783192396164, + 0.015503749251365662, + 0.014558619819581509, + 0.004347691312432289, + 0.004921548068523407, + 0.006659587379544973, + 0.00441946042701602, + 0.004743943922221661, + 0.010963981039822102, + 0.0038892629090696573, + 0.004382440820336342, + 0.006616541184484959, + 0.01330643892288208, + 0.00884312018752098, + 0.010399291291832924, + 0.007826177403330803, + 0.013973472639918327, + 0.006483911536633968, + 0.005554533563554287, + 0.005247754510492086, + 0.004808003548532724, + 0.004779106471687555, + 0.005805226042866707, + 0.005871088244020939, + 0.006913468241691589, + 0.005105063784867525, + 0.005614664405584335, + 0.005034714937210083, + 0.003892567241564393, + 0.007231712341308594, + 0.0055048163048923016, + 0.004264292307198048, + 0.006367036607116461, + 0.0049871825613081455, + 0.009249879978597164, + 0.0131689989939332, + 0.009799625724554062, + 0.02002839744091034, + 0.011448867619037628, + 0.013414346612989902, + 0.023908590897917747, + 0.02695128135383129, + 0.013745070435106754, + 0.0072985561564564705, + 0.005863490514457226, + 0.008314020931720734, + 0.007770562078803778, + 0.006721983198076487, + 0.006547544617205858, + 0.0070863221772015095, + 0.005593062378466129, + 0.019852789118885994, + 0.01911189965903759, + 0.024931427091360092, + 0.016816379502415657, + 0.014260422438383102, + 0.018661201000213623, + 0.013869262300431728, + 0.01598064973950386, + 0.03535476326942444, + 0.04188574105501175, + 0.051174893975257874, + 0.023789558559656143, + 0.04888303950428963, + 0.03937545791268349, + 0.031799815595149994, + 0.024794932454824448, + 0.008150799199938774, + 0.014248937368392944, + 0.010772103443741798, + 0.011145251803100109, + 0.006132683251053095, + 0.00510967755690217, + 0.006132226437330246, + 0.0038580975960940123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.07001003623008728 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176652491092682 + ], + "opposite_logprob_tokens": [ + -4.018176555633545 + ], + "logprob_reward": -0.018176652491092682, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176652491092682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08959322422742844, + 0.11123180389404297, + 0.18339936435222626, + 0.09561078250408173, + 0.07972726970911026, + 0.03804554417729378, + 0.13593870401382446, + 0.09184031933546066, + 0.092705138027668, + 0.08190782368183136 + ], + "credit_map": [ + 0.009395931847393513, + 0.008342073298990726, + 0.011610191315412521, + 0.011366627179086208, + 0.006741284392774105, + 0.014995014294981956, + 0.018630139529705048, + 0.008511965163052082, + 0.01254053320735693, + 0.00942297838628292, + 0.012945888563990593, + 0.012356595136225224, + 0.011072738096117973, + 0.02039862610399723, + 0.013419808819890022, + 0.01907464861869812, + 0.018072322010993958, + 0.026929764077067375, + 0.022836776450276375, + 0.0238527562469244, + 0.03818218782544136, + 0.018072757869958878, + 0.016616348177194595, + 0.018836455419659615, + 0.008958053775131702, + 0.008911947719752789, + 0.01664884015917778, + 0.010011936537921429, + 0.011243934743106365, + 0.021409718319773674, + 0.007683414500206709, + 0.010742942802608013, + 0.0076574417762458324, + 0.006031370256096125, + 0.011212512850761414, + 0.009192765690386295, + 0.006130711175501347, + 0.012400361709296703, + 0.012051157653331757, + 0.01505095511674881, + 0.006011576391756535, + 0.0029510254971683025, + 0.006442815996706486, + 0.002789292251691222, + 0.005078177899122238, + 0.007417353335767984, + 0.004049632698297501, + 0.0033056745305657387, + 0.018835360184311867, + 0.009250801056623459, + 0.015602379105985165, + 0.017293022945523262, + 0.011887339875102043, + 0.03277849778532982, + 0.014612141065299511, + 0.015679167583584785, + 0.015238860622048378, + 0.011163736693561077, + 0.018189704045653343, + 0.013361478224396706, + 0.006114810239523649, + 0.010384371504187584, + 0.01057870127260685, + 0.006808661390095949, + 0.011445430107414722, + 0.012657491490244865, + 0.01621975004673004, + 0.00847464520484209, + 0.013327484019100666, + 0.012369280681014061, + 0.00952258612960577, + 0.008688487112522125, + 0.012535249814391136, + 0.011887800879776478, + 0.017443804070353508, + 0.013391944579780102, + 0.006396648474037647, + 0.007001955062150955, + 0.007411875296384096, + 0.0058385576121509075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.018176652491092682 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018175246194005013 + ], + "opposite_logprob_tokens": [ + -4.01817512512207 + ], + "logprob_reward": -0.018175246194005013, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018175246194005013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12488291412591934, + 0.03182782232761383, + 0.043633513152599335, + 0.07053352892398834, + 0.043676406145095825, + 0.04442930966615677, + 0.10024133324623108, + 0.07461261004209518, + 0.16941219568252563, + 0.2967503070831299 + ], + "credit_map": [ + 0.012421081773936749, + 0.010901754721999168, + 0.015933843329548836, + 0.013899859972298145, + 0.012238848954439163, + 0.014606024138629436, + 0.028856931254267693, + 0.016024567186832428, + 0.0029170745983719826, + 0.0058375876396894455, + 0.004279625602066517, + 0.005388488527387381, + 0.003195682540535927, + 0.00348403281532228, + 0.003355841850861907, + 0.0033694878220558167, + 0.004142502322793007, + 0.005788110662251711, + 0.006236894987523556, + 0.006825354415923357, + 0.005292275454849005, + 0.005259486380964518, + 0.005231678020209074, + 0.004857211373746395, + 0.008769465610384941, + 0.0058589838445186615, + 0.01865801215171814, + 0.0059972964227199554, + 0.0036002439446747303, + 0.014734810218214989, + 0.0043219905346632, + 0.008592727594077587, + 0.005928764119744301, + 0.0024598692543804646, + 0.006109097506850958, + 0.0035738174337893724, + 0.003028696170076728, + 0.009558209218084812, + 0.004590592347085476, + 0.008427360095083714, + 0.008372648619115353, + 0.0041569857858121395, + 0.0062284916639328, + 0.004143456928431988, + 0.0052884649485349655, + 0.006975517608225346, + 0.0037524716462939978, + 0.005511271767318249, + 0.013284904882311821, + 0.007987634278833866, + 0.011278539896011353, + 0.014064785093069077, + 0.008285348303616047, + 0.02156001515686512, + 0.01103232428431511, + 0.012747782282531261, + 0.011276420205831528, + 0.00853449385613203, + 0.013916491530835629, + 0.010499585419893265, + 0.006224576849490404, + 0.007774596568197012, + 0.006492980755865574, + 0.009893465787172318, + 0.022244708612561226, + 0.022550547495484352, + 0.031414229422807693, + 0.015181017108261585, + 0.02351120114326477, + 0.02329256758093834, + 0.01689225435256958, + 0.01432567834854126, + 0.041266243904829025, + 0.0608283206820488, + 0.0401051864027977, + 0.06148039177060127, + 0.023251889273524284, + 0.02149338647723198, + 0.0274774432182312, + 0.020847441628575325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.018175246194005013 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, taken from a slight overhead angle, focuses on a chubby orange tabby cat nestled in an interior space with polished light-brown wood floors. Warm afternoon sunlight slants in from outside the frame, casting a bright patch of light on the floor. Within this patch of light, the cat, arching its chubby body, begins a series of rapid, clumsy spins, pivoting on its front legs. Concentrated in pursuit of its long, furry tail, its movements appear uncoordinated due to its sleek shape, each spin imbued with a comical wobbling quality. Its short, sturdy limbs struggle to maintain balance, yet its claws repeatedly reach out in vain. The sunlight illuminates its fluffy, soft fur, casting a golden halo around the edges. Its green eyes, fixed intently on the tip of its tail, which it can't grasp, exude a sense of both persistence and innocence. The warm tones of the entire scene create a deeply warm, relaxed, and playful atmosphere.", + "seed": 3045, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_10/index_6322_seed_3045_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06483579571358859, + "vlm_logprob": -0.06483579571358859, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.06483579571358859, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03540075570344925 + ], + "opposite_logprob_tokens": [ + -3.410400867462158 + ], + "logprob_reward": -0.03540075570344925, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03540075570344925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22130261361598969, + 0.45252200961112976, + 0.02513422630727291, + 0.020942943170666695, + 0.017069904133677483, + 0.030114464461803436, + 0.03621247410774231, + 0.053394418209791183, + 0.10417202115058899, + 0.039135001599788666 + ], + "credit_map": [ + 0.03347310796380043, + 0.025663770735263824, + 0.03182339668273926, + 0.030070584267377853, + 0.019619418308138847, + 0.02560354582965374, + 0.027131278067827225, + 0.027917470782995224, + 0.140297994017601, + 0.01669849455356598, + 0.1860959678888321, + 0.041262559592723846, + 0.01649591326713562, + 0.014709251001477242, + 0.01876019686460495, + 0.018201561644673347, + 0.003003999125212431, + 0.0023763643112033606, + 0.003101328620687127, + 0.005555153824388981, + 0.002094535157084465, + 0.0019021311309188604, + 0.004465386271476746, + 0.002635322045534849, + 0.002689399290829897, + 0.0017940570833161473, + 0.0030345686245709658, + 0.004160851240158081, + 0.001732932054437697, + 0.0016389291267842054, + 0.0033209428656846285, + 0.0025712584611028433, + 0.0019848416559398174, + 0.0016581997042521834, + 0.0024785506539046764, + 0.003123943228274584, + 0.0012110430980101228, + 0.0012807190651074052, + 0.003308187937363982, + 0.0020244175102561712, + 0.003970090299844742, + 0.0021830161567777395, + 0.00391142163425684, + 0.0052346824668347836, + 0.0030604859348386526, + 0.002641320461407304, + 0.004472849424928427, + 0.004640594124794006, + 0.004624329973012209, + 0.004061444662511349, + 0.005716788582503796, + 0.004404671024531126, + 0.003562493249773979, + 0.004849132616072893, + 0.0051098535768687725, + 0.003883753903210163, + 0.008783828467130661, + 0.0044342936016619205, + 0.012550760060548782, + 0.005470470059663057, + 0.005944765172898769, + 0.0040873074904084206, + 0.006669137626886368, + 0.005453846883028746, + 0.010389142669737339, + 0.016413992270827293, + 0.013656008057296276, + 0.015603244304656982, + 0.01438585389405489, + 0.00879149604588747, + 0.014451242983341217, + 0.0104810306802392, + 0.004240746144205332, + 0.004212278872728348, + 0.0045646363869309425, + 0.00656379759311676, + 0.004047352820634842, + 0.0026637485716491938, + 0.007661617826670408, + 0.005180817563086748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.03540075570344925 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741280667483807 + ], + "opposite_logprob_tokens": [ + -5.006741046905518 + ], + "logprob_reward": -0.006741280667483807, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741280667483807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09486300498247147, + 0.19379521906375885, + 0.05447685718536377, + 0.15257668495178223, + 0.045019589364528656, + 0.051254115998744965, + 0.09322056919336319, + 0.19578370451927185, + 0.04381084814667702, + 0.07519939541816711 + ], + "credit_map": [ + 0.015164582058787346, + 0.012614027597010136, + 0.01069159060716629, + 0.01619364321231842, + 0.00859933253377676, + 0.007130172569304705, + 0.011705798096954823, + 0.012763856910169125, + 0.05441256985068321, + 0.011566360481083393, + 0.0756528452038765, + 0.01763470470905304, + 0.007829392328858376, + 0.006722071673721075, + 0.012318761087954044, + 0.007658522576093674, + 0.007941859774291515, + 0.0076433634385466576, + 0.006944079417735338, + 0.008745272643864155, + 0.0046106851659715176, + 0.005237444303929806, + 0.007689958438277245, + 0.005664194002747536, + 0.034201592206954956, + 0.01172486413270235, + 0.016610171645879745, + 0.019136864691972733, + 0.00815549585968256, + 0.011407299898564816, + 0.013840622268617153, + 0.03749977797269821, + 0.006864661816507578, + 0.0033550288062542677, + 0.011135845445096493, + 0.004648939706385136, + 0.0034344184678047895, + 0.005847247317433357, + 0.004218028858304024, + 0.0055154175497591496, + 0.008813193067908287, + 0.004324496723711491, + 0.010216833092272282, + 0.006729365326464176, + 0.003869424806907773, + 0.004278283566236496, + 0.005854576826095581, + 0.007167940493673086, + 0.021601025015115738, + 0.007663872092962265, + 0.02436346374452114, + 0.008594772778451443, + 0.00534019572660327, + 0.0052339318208396435, + 0.009214349091053009, + 0.011208957061171532, + 0.032872892916202545, + 0.0177159421145916, + 0.03451668843626976, + 0.01107990462332964, + 0.021511932834982872, + 0.02636154554784298, + 0.022705385461449623, + 0.029019411653280258, + 0.004763494711369276, + 0.005343098659068346, + 0.005917152389883995, + 0.005706368945538998, + 0.00581999821588397, + 0.0038790518883615732, + 0.007023571524769068, + 0.005358110647648573, + 0.008845326490700245, + 0.008996640332043171, + 0.009042143821716309, + 0.007502108346670866, + 0.009688081219792366, + 0.006534162908792496, + 0.012057827785611153, + 0.012533101253211498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.006741280667483807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25208303332328796 + ], + "opposite_logprob_tokens": [ + -1.5020830631256104 + ], + "logprob_reward": -0.25208303332328796, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25208303332328796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1089932918548584, + 0.3224714994430542, + 0.04371929541230202, + 0.03975808992981911, + 0.01651066355407238, + 0.0956253930926323, + 0.12971679866313934, + 0.10757875442504883, + 0.040041014552116394, + 0.09558519721031189 + ], + "credit_map": [ + 0.017917491495609283, + 0.010671377182006836, + 0.017103299498558044, + 0.013137873262166977, + 0.010884115472435951, + 0.01283236499875784, + 0.012345468625426292, + 0.014101295731961727, + 0.09229547530412674, + 0.011688976548612118, + 0.13471049070358276, + 0.03108077682554722, + 0.012804773636162281, + 0.009308701381087303, + 0.01341851893812418, + 0.01716379076242447, + 0.0040487265214324, + 0.003895585425198078, + 0.0056106350384652615, + 0.008420944213867188, + 0.00511609623208642, + 0.004483269061893225, + 0.006790513638406992, + 0.005353523883968592, + 0.006563892122358084, + 0.003203270025551319, + 0.0059000467881560326, + 0.006316060200333595, + 0.003575071459636092, + 0.0023119617253541946, + 0.005511515308171511, + 0.006376274395734072, + 0.0021148500964045525, + 0.0016393369296565652, + 0.001928989659063518, + 0.002558446489274502, + 0.0012363565620034933, + 0.0019762502051889896, + 0.0028498833999037743, + 0.0022065499797463417, + 0.01470904890447855, + 0.006566185504198074, + 0.0166489090770483, + 0.014838787727057934, + 0.008817236870527267, + 0.006745683960616589, + 0.014435762539505959, + 0.01286377850919962, + 0.03306645154953003, + 0.016050219535827637, + 0.035812657326459885, + 0.007992321625351906, + 0.008400954306125641, + 0.00772447744384408, + 0.009237036108970642, + 0.01143267098814249, + 0.01490231603384018, + 0.009516210295259953, + 0.01876629702746868, + 0.008229470811784267, + 0.013107326813042164, + 0.014991089701652527, + 0.013308698311448097, + 0.014757353812456131, + 0.00540969567373395, + 0.004380084108561277, + 0.005022997967898846, + 0.003017942886799574, + 0.0048256185837090015, + 0.00638602627441287, + 0.005931807681918144, + 0.005066841375082731, + 0.009872193448245525, + 0.011790547519922256, + 0.010912737809121609, + 0.011608975939452648, + 0.010094030760228634, + 0.012175768613815308, + 0.017331022769212723, + 0.011799921281635761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.25208303332328796 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026339229196310043 + ], + "opposite_logprob_tokens": [ + -3.651339292526245 + ], + "logprob_reward": -0.026339229196310043, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026339229196310043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051069024950265884, + 0.6369374394416809, + 0.04638995602726936, + 0.01559114083647728, + 0.029235869646072388, + 0.027935165911912918, + 0.09288956969976425, + 0.0519292876124382, + 0.024473849684000015, + 0.02354871854186058 + ], + "credit_map": [ + 0.007264730520546436, + 0.008175607770681381, + 0.006149028427898884, + 0.007836548611521721, + 0.004744838923215866, + 0.0045916372910141945, + 0.005936004687100649, + 0.006370630115270615, + 0.20802466571331024, + 0.020103679969906807, + 0.2727425992488861, + 0.057019274681806564, + 0.0232478566467762, + 0.016382785513997078, + 0.022065259516239166, + 0.017351320013403893, + 0.005296553950756788, + 0.004155477974563837, + 0.006230952683836222, + 0.009862279519438744, + 0.003994111903011799, + 0.0028555963654071093, + 0.00947881955653429, + 0.00451616570353508, + 0.002688124543055892, + 0.0012371842749416828, + 0.0021364991553127766, + 0.0022518516052514315, + 0.0013525086687877774, + 0.0013336599804461002, + 0.0019265789305791259, + 0.0026647320482879877, + 0.0031273833010345697, + 0.0024946280755102634, + 0.003867305116727948, + 0.005680427420884371, + 0.002120457822456956, + 0.002039602492004633, + 0.005890755448490381, + 0.004015308804810047, + 0.0038132858462631702, + 0.0023828332778066397, + 0.003718276973813772, + 0.004353134427219629, + 0.003027559258043766, + 0.0022047085221856833, + 0.00427827425301075, + 0.004157091956585646, + 0.020027030259370804, + 0.01375185139477253, + 0.022983578965067863, + 0.006670608185231686, + 0.007151133380830288, + 0.006014370825141668, + 0.007454332429915667, + 0.008836664259433746, + 0.009661227464675903, + 0.003738625207915902, + 0.009409419260919094, + 0.0042534624226391315, + 0.004738876596093178, + 0.004150770138949156, + 0.009760009124875069, + 0.006216895766556263, + 0.0026899315416812897, + 0.003869627369567752, + 0.0025976889301091433, + 0.003061773721128702, + 0.003781606676056981, + 0.001908490783534944, + 0.003509405069053173, + 0.003055326407775283, + 0.003111056750640273, + 0.0024659077171236277, + 0.0035862650256603956, + 0.003558630822226405, + 0.002864264417439699, + 0.001530380453914404, + 0.0038410115521401167, + 0.0025912022683769464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.026339229196310043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003614679677411914 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003614679677411914, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614679677411914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10737193375825882, + 0.24999567866325378, + 0.11434949189424515, + 0.02801469899713993, + 0.2086823582649231, + 0.05081671103835106, + 0.05340242385864258, + 0.09709644317626953, + 0.029737021774053574, + 0.0605333149433136 + ], + "credit_map": [ + 0.014639975503087044, + 0.014300277456641197, + 0.01730937883257866, + 0.01677643693983555, + 0.008705570362508297, + 0.008142473176121712, + 0.014428628608584404, + 0.01306919101625681, + 0.07124023884534836, + 0.008255704306066036, + 0.08832208067178726, + 0.0229698084294796, + 0.012546151876449585, + 0.01129617914557457, + 0.021876275539398193, + 0.013489240780472755, + 0.016210779547691345, + 0.013142317533493042, + 0.019440749660134315, + 0.012961973436176777, + 0.009272392839193344, + 0.014115868136286736, + 0.01113268081098795, + 0.01807272806763649, + 0.00366664188914001, + 0.001861182157881558, + 0.0044695730321109295, + 0.004768397659063339, + 0.002054385608062148, + 0.0024351098109036684, + 0.004010478034615517, + 0.004748929757624865, + 0.028626268729567528, + 0.01916000433266163, + 0.034870728850364685, + 0.02937167137861252, + 0.012841827236115932, + 0.02293810062110424, + 0.028324535116553307, + 0.03254921734333038, + 0.006896254140883684, + 0.004326603841036558, + 0.007648214232176542, + 0.006084132939577103, + 0.0046720136888325214, + 0.007206901907920837, + 0.006587364245206118, + 0.007395225577056408, + 0.010055175982415676, + 0.005089289508759975, + 0.01053022500127554, + 0.00510579627007246, + 0.004800323862582445, + 0.005197049118578434, + 0.006164225749671459, + 0.006460339762270451, + 0.019871359691023827, + 0.006478454917669296, + 0.019322721287608147, + 0.010207054205238819, + 0.0076742712408304214, + 0.008227809332311153, + 0.012887132354080677, + 0.012427637353539467, + 0.0031394970137625933, + 0.004378965590149164, + 0.00412617065012455, + 0.0035654674284160137, + 0.004020331893116236, + 0.0033430454786866903, + 0.0038943218532949686, + 0.0032692200038582087, + 0.007091349922120571, + 0.00684390077367425, + 0.00649686437100172, + 0.006066763307899237, + 0.008196978829801083, + 0.004480871371924877, + 0.012506281025707722, + 0.00885030534118414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.003614679677411914 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, taken from a slight overhead angle, focuses on a chubby orange tabby cat nestled in an interior space with polished light-brown wood floors. Warm afternoon sunlight slants in from outside the frame, casting a bright patch of light on the floor. Within this patch of light, the cat, arching its chubby body, begins a series of rapid, clumsy spins, pivoting on its front legs. Concentrated in pursuit of its long, furry tail, its movements appear uncoordinated due to its sleek shape, each spin imbued with a comical wobbling quality. Its short, sturdy limbs struggle to maintain balance, yet its claws repeatedly reach out in vain. The sunlight illuminates its fluffy, soft fur, casting a golden halo around the edges. Its green eyes, fixed intently on the tip of its tail, which it can't grasp, exude a sense of both persistence and innocence. The warm tones of the entire scene create a deeply warm, relaxed, and playful atmosphere.", + "seed": 3046, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_10/index_6322_seed_3046_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07267715157940983, + "vlm_logprob": -0.07267715157940983, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.07267715157940983, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2031070590019226 + ], + "opposite_logprob_tokens": [ + -1.7031071186065674 + ], + "logprob_reward": -0.2031070590019226, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2031070590019226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2195255607366562, + 0.043949488550424576, + 0.02499474585056305, + 0.15094642341136932, + 0.07994779199361801, + 0.015829337760806084, + 0.11866705864667892, + 0.13084101676940918, + 0.18403173983097076, + 0.031266748905181885 + ], + "credit_map": [ + 0.026525182649493217, + 0.03514456748962402, + 0.026321910321712494, + 0.020509235560894012, + 0.02843383513391018, + 0.02231702394783497, + 0.029614398255944252, + 0.03065941296517849, + 0.0033994168043136597, + 0.007286548148840666, + 0.01014502253383398, + 0.010375776328146458, + 0.0033773111645132303, + 0.0029789344407618046, + 0.003624922363087535, + 0.002761559095233679, + 0.0024125566706061363, + 0.00634823739528656, + 0.0018382156267762184, + 0.0036813016049563885, + 0.0035781527403742075, + 0.0017357927281409502, + 0.0035250785294920206, + 0.0018754133488982916, + 0.014972941018640995, + 0.02309834212064743, + 0.014761583879590034, + 0.02150692790746689, + 0.02818010002374649, + 0.010233721695840359, + 0.026318151503801346, + 0.011874668300151825, + 0.010722353123128414, + 0.00588771840557456, + 0.0081831905990839, + 0.01297099981456995, + 0.009211420081555843, + 0.007082263007760048, + 0.00925634615123272, + 0.016633516177535057, + 0.0017213185783475637, + 0.002045086119323969, + 0.0013580105733126402, + 0.0021528578363358974, + 0.0022100748028606176, + 0.0023066948633641005, + 0.002312487456947565, + 0.001722809742204845, + 0.013919965364038944, + 0.02406713180243969, + 0.010707256384193897, + 0.015183055773377419, + 0.014743386767804623, + 0.01019209623336792, + 0.020260116085410118, + 0.009594063274562359, + 0.017176464200019836, + 0.021155135706067085, + 0.012984206900000572, + 0.017180999740958214, + 0.01618020236492157, + 0.010830164887011051, + 0.019795799627900124, + 0.01553806196898222, + 0.01808415912091732, + 0.01705024763941765, + 0.024846011772751808, + 0.01802746392786503, + 0.02505282685160637, + 0.0151951489970088, + 0.024517487734556198, + 0.04125840216875076, + 0.002805836498737335, + 0.0031621260568499565, + 0.003026543417945504, + 0.0031143000815063715, + 0.005335857626050711, + 0.004305950831621885, + 0.006995201576501131, + 0.002520934911444783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.2031070590019226 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642185289412737 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004642185289412737, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642185289412737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07634645700454712, + 0.2679317593574524, + 0.06421912461519241, + 0.05960964411497116, + 0.0802776962518692, + 0.047710493206977844, + 0.1152084469795227, + 0.042643219232559204, + 0.21559734642505646, + 0.030455803498625755 + ], + "credit_map": [ + 0.011435134336352348, + 0.009813747368752956, + 0.011382941156625748, + 0.009198585525155067, + 0.007445715833455324, + 0.006567848846316338, + 0.008533991873264313, + 0.01196849625557661, + 0.02403107099235058, + 0.015088988468050957, + 0.10587478429079056, + 0.08980491012334824, + 0.009120865724980831, + 0.006486828438937664, + 0.011683222837746143, + 0.0058410875499248505, + 0.006175707094371319, + 0.007765099871903658, + 0.006049490533769131, + 0.009706580080091953, + 0.009169548749923706, + 0.009027981199324131, + 0.007991098798811436, + 0.008333618752658367, + 0.007107059471309185, + 0.008183636702597141, + 0.006233955733478069, + 0.008251547813415527, + 0.01002726424485445, + 0.005218703765422106, + 0.00898302998393774, + 0.005604446399956942, + 0.00775730749592185, + 0.005220239982008934, + 0.02410798706114292, + 0.015079114586114883, + 0.005975514650344849, + 0.005198987200856209, + 0.006787126883864403, + 0.010151421651244164, + 0.004823924973607063, + 0.008211750537157059, + 0.003862538607791066, + 0.0047460077330470085, + 0.007025259546935558, + 0.005603860132396221, + 0.007352747023105621, + 0.006084405351430178, + 0.012359567917883396, + 0.01942315697669983, + 0.01278198417276144, + 0.017819853499531746, + 0.012933459132909775, + 0.014797588810324669, + 0.014472068287432194, + 0.010620773769915104, + 0.00524466997012496, + 0.005664799362421036, + 0.004590930417180061, + 0.0054797399789094925, + 0.005585499573498964, + 0.00397584680467844, + 0.005975146312266588, + 0.006126584950834513, + 0.021944768726825714, + 0.00546340923756361, + 0.05476599931716919, + 0.020089175552129745, + 0.004025526810437441, + 0.015591057948768139, + 0.018213652074337006, + 0.07550376653671265, + 0.0037837978452444077, + 0.004120837897062302, + 0.00252967095002532, + 0.003407962154597044, + 0.0043502077460289, + 0.004330006893724203, + 0.00481379684060812, + 0.0031195227056741714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.004642185289412737 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270400732755661 + ], + "opposite_logprob_tokens": [ + -2.127040147781372 + ], + "logprob_reward": -0.1270400732755661, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270400732755661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18395155668258667, + 0.1156311184167862, + 0.0815037339925766, + 0.1198134496808052, + 0.11514682322740555, + 0.02740517072379589, + 0.04028565064072609, + 0.06540914624929428, + 0.22872409224510193, + 0.02212926186621189 + ], + "credit_map": [ + 0.016498658806085587, + 0.02508513256907463, + 0.030206646770238876, + 0.021680234000086784, + 0.019291702657938004, + 0.015358799137175083, + 0.026497622951865196, + 0.029332755133509636, + 0.010483883321285248, + 0.01587338000535965, + 0.02278054505586624, + 0.026841619983315468, + 0.01347489282488823, + 0.008308631367981434, + 0.011826744303107262, + 0.006041421089321375, + 0.008592771366238594, + 0.012955890968441963, + 0.005806418135762215, + 0.010884297080338001, + 0.013029300607740879, + 0.008284485898911953, + 0.016116293147206306, + 0.005834277719259262, + 0.012136291712522507, + 0.01876307651400566, + 0.01264230627566576, + 0.017105260863900185, + 0.021378565579652786, + 0.008364448323845863, + 0.019759416580200195, + 0.009664079174399376, + 0.011883500963449478, + 0.010080712847411633, + 0.015177218243479729, + 0.01702786050736904, + 0.010179480537772179, + 0.0131229842081666, + 0.017383314669132233, + 0.02029174007475376, + 0.0034259450621902943, + 0.004008779767900705, + 0.0024535367265343666, + 0.0029826597310602665, + 0.004462843295186758, + 0.003549061017110944, + 0.0032557693775743246, + 0.00326657691039145, + 0.004315264988690615, + 0.006812931504100561, + 0.004308377392590046, + 0.005496231373399496, + 0.005146427545696497, + 0.003428194671869278, + 0.006602016277611256, + 0.004176207818090916, + 0.007597789634019136, + 0.00945421401411295, + 0.006371095776557922, + 0.00845429114997387, + 0.00845314934849739, + 0.006276219617575407, + 0.00832275953143835, + 0.010479624383151531, + 0.023293379694223404, + 0.009675727225840092, + 0.052698444575071335, + 0.021178238093852997, + 0.010648729279637337, + 0.016234412789344788, + 0.01973508857190609, + 0.07526005059480667, + 0.0024561607278883457, + 0.0025200089439749718, + 0.0023589497432112694, + 0.0019889501854777336, + 0.0032553498167544603, + 0.002937972778454423, + 0.004124640487134457, + 0.002487230347469449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.1270400732755661 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063205897808075 + ], + "opposite_logprob_tokens": [ + -4.141063213348389 + ], + "logprob_reward": -0.016063205897808075, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063205897808075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04596925526857376, + 0.5492698550224304, + 0.03142222762107849, + 0.025567490607500076, + 0.049780312925577164, + 0.013914162293076515, + 0.020086253061890602, + 0.02595016174018383, + 0.21145467460155487, + 0.026585495099425316 + ], + "credit_map": [ + 0.0061312271282076836, + 0.006041116081178188, + 0.0036013179924339056, + 0.004634804092347622, + 0.00712843332439661, + 0.00504818931221962, + 0.005301978439092636, + 0.008082189597189426, + 0.04939688742160797, + 0.02505040355026722, + 0.22570419311523438, + 0.19192242622375488, + 0.01335976179689169, + 0.010932298377156258, + 0.02129737287759781, + 0.011606527492403984, + 0.003249258268624544, + 0.005168376490473747, + 0.0028264622669667006, + 0.003948248457163572, + 0.003799578407779336, + 0.004330303054302931, + 0.004974065348505974, + 0.003125937655568123, + 0.003461908083409071, + 0.0029534571804106236, + 0.0029185398016124964, + 0.002932655857875943, + 0.004493444226682186, + 0.00178534013684839, + 0.004114537034183741, + 0.0029076095670461655, + 0.005412720143795013, + 0.003929929342120886, + 0.006077940575778484, + 0.006424425169825554, + 0.006865305360406637, + 0.007275644224137068, + 0.007506044115871191, + 0.006288307718932629, + 0.0013559212675318122, + 0.001878429320640862, + 0.0011692673433572054, + 0.0017042499966919422, + 0.0019948387052863836, + 0.0017816369654610753, + 0.002184598706662655, + 0.0018452199874445796, + 0.0021644150838255882, + 0.003515311051160097, + 0.002112765097990632, + 0.002747219055891037, + 0.0021729706786572933, + 0.0022814616095274687, + 0.0031643195543438196, + 0.0019277897663414478, + 0.00353849190287292, + 0.00395863875746727, + 0.0030658820178359747, + 0.003179591614753008, + 0.0032139604445546865, + 0.0024817639496177435, + 0.003453123616054654, + 0.0030587080400437117, + 0.021092398092150688, + 0.00466232281178236, + 0.05423156917095184, + 0.01952694170176983, + 0.004365524742752314, + 0.015301082283258438, + 0.017956122756004333, + 0.07431869953870773, + 0.003380178939551115, + 0.003386132651939988, + 0.002225517760962248, + 0.0025436473079025745, + 0.0034742234274744987, + 0.004699438810348511, + 0.004051710944622755, + 0.002824646420776844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.016063205897808075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chubby orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set indoors with polished light-brown wood floors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cat perform rapid, clumsy spins while arching its body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a static medium shot taken from a slight overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is warm afternoon sunlight visible casting a bright patch of light on the floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533234432339668 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.012533234432339668, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533234432339668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1524585485458374, + 0.10791505128145218, + 0.052349865436553955, + 0.24666595458984375, + 0.03496376797556877, + 0.05133714899420738, + 0.0519140288233757, + 0.03247286006808281, + 0.2445428967475891, + 0.025379886850714684 + ], + "credit_map": [ + 0.019112132489681244, + 0.015581781975924969, + 0.02724308893084526, + 0.015519923530519009, + 0.014389867894351482, + 0.012432243674993515, + 0.015983417630195618, + 0.03219607472419739, + 0.01622786559164524, + 0.008467930369079113, + 0.03075215592980385, + 0.02120187133550644, + 0.007333606947213411, + 0.006777338217943907, + 0.009295530617237091, + 0.00785875041037798, + 0.0054980251006782055, + 0.007554813288152218, + 0.005185253452509642, + 0.008842194452881813, + 0.00658124266192317, + 0.008302140980958939, + 0.005380307789891958, + 0.0050058867782354355, + 0.03158324211835861, + 0.016334405168890953, + 0.0349457785487175, + 0.027072975412011147, + 0.02817346341907978, + 0.03612133488059044, + 0.028942538425326347, + 0.04349221661686897, + 0.0034500358160585165, + 0.0034191154409199953, + 0.005790194496512413, + 0.005048982799053192, + 0.004105565138161182, + 0.004059370141476393, + 0.00407756632193923, + 0.0050129359588027, + 0.00578782195225358, + 0.00720039801672101, + 0.003837033174932003, + 0.005071575753390789, + 0.009271377697587013, + 0.007675390224903822, + 0.007079818286001682, + 0.005413732957094908, + 0.006350982002913952, + 0.008822683244943619, + 0.0057061221450567245, + 0.007770069409161806, + 0.005137829575687647, + 0.005815105512738228, + 0.006836663465946913, + 0.005474573466926813, + 0.00366991781629622, + 0.0038696499541401863, + 0.0031425978522747755, + 0.0034662552643567324, + 0.00480104424059391, + 0.0037164255045354366, + 0.004985095001757145, + 0.004821872338652611, + 0.024695757776498795, + 0.006486658472567797, + 0.06264879554510117, + 0.02335810475051403, + 0.0044494555331766605, + 0.017332930117845535, + 0.020504726096987724, + 0.08506646007299423, + 0.003492872230708599, + 0.0032732870895415545, + 0.0020267197396606207, + 0.0027008862234652042, + 0.002843318972736597, + 0.004344351124018431, + 0.003804634790867567, + 0.0028938152827322483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.012533234432339668 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_11/index_tensor([3592, 3592])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_11/index_tensor([3592, 3592])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..27448e46e6e64bb641d26e5a763f5a659af79312 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_11/index_tensor([3592, 3592])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, tightly packed portrait-style close-up centers on the chest area of an unseen figure, where ethereal streams of fiery purple energy rapidly gather and weave into a shimmering, translucent bubble of magical energy—the Banshee's Veil. Its surface, a delicate crystalline grid, hums with contained power, casting a soft purple glow on the invisible armor beneath. Just as the sphere reaches its full, perfect form, a burst of searing crimson energy radiates from off-screen, violently colliding with the barrier and triggering a breathtaking chain reaction of deep purple light that ripples across the shield's surface in concentric waves, causing the entire structure to buckle and twist under the immense force. As the shield steadily and deliberately dissipates, fragmenting into a series of shimmering, fading points of light, the camera smoothly tilts upward, rising above the ornate dark steel armor to reveal the resolute face of a female warrior. Her face, framed by wisps of sweat-stained, raven-black hair, freed from its tight braid, wore a mask of concentration; her sharp, high cheekbones were illuminated by the residual purple light, her jaw set, her sharp grey eyes narrowed, unblinking, staring resolutely ahead, creating a tense and heroic atmosphere of a narrow victory.", + "seed": 12275, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_11/index_3592_seed_12275_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06684616226702929, + "vlm_logprob": -0.06684616226702929, + "vlm_yes_no_margin": 3.174999976158142, + "vlm_token_credit": -0.06684616226702929, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13225796818733215 + ], + "opposite_logprob_tokens": [ + -2.1322579383850098 + ], + "logprob_reward": -0.13225796818733215, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13225796818733215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07523076981306076, + 0.051381632685661316, + 0.15717045962810516, + 0.09808970242738724, + 0.2134031504392624, + 0.18056295812129974, + 0.17383287847042084, + 0.014423081651329994, + 0.018304046243429184, + 0.0176013745367527 + ], + "credit_map": [ + 0.007130925543606281, + 0.017943229526281357, + 0.007887194864451885, + 0.007019687443971634, + 0.012809273786842823, + 0.008775359950959682, + 0.006982847116887569, + 0.006682252511382103, + 0.004464052151888609, + 0.0049864924512803555, + 0.00575855141505599, + 0.007508445996791124, + 0.008239593356847763, + 0.006456468719989061, + 0.007950819097459316, + 0.0060172067023813725, + 0.009931068867444992, + 0.013596145436167717, + 0.03178328275680542, + 0.06224311143159866, + 0.008208872750401497, + 0.007086750585585833, + 0.007424077484756708, + 0.01689714565873146, + 0.00915538240224123, + 0.03872774913907051, + 0.014606446959078312, + 0.008969845250248909, + 0.008760401979088783, + 0.006972181145101786, + 0.003949502483010292, + 0.006948194466531277, + 0.022100256755948067, + 0.021847279742360115, + 0.024111364036798477, + 0.02448119968175888, + 0.02912214770913124, + 0.03310517594218254, + 0.03470991179347038, + 0.023925811052322388, + 0.01945827528834343, + 0.019880278035998344, + 0.02193121798336506, + 0.04798443615436554, + 0.017961183562874794, + 0.019893795251846313, + 0.017899207770824432, + 0.01555455569177866, + 0.022373385727405548, + 0.012452500872313976, + 0.02394947037100792, + 0.029842304065823555, + 0.01844550482928753, + 0.02327990159392357, + 0.02598855085670948, + 0.01750124990940094, + 0.0018145448993891478, + 0.001734851161018014, + 0.001711373683065176, + 0.002475651679560542, + 0.0016373805701732635, + 0.0016961252549663186, + 0.0016436640871688724, + 0.001709490898065269, + 0.0021107348147779703, + 0.0010839850874617696, + 0.00231632380746305, + 0.0018470055656507611, + 0.0031144132371991873, + 0.0020426083356142044, + 0.0018549085361883044, + 0.00393406767398119, + 0.0019011357799172401, + 0.0019413245609030128, + 0.001966073177754879, + 0.0022156760096549988, + 0.003570531029254198, + 0.0017692241817712784, + 0.0024124488700181246, + 0.00182496034540236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.13225796818733215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029839513823390007 + ], + "opposite_logprob_tokens": [ + -3.529839515686035 + ], + "logprob_reward": -0.029839513823390007, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029839513823390007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08617933839559555, + 0.05112558603286743, + 0.2990500330924988, + 0.052300672978162766, + 0.09474506229162216, + 0.05129000544548035, + 0.057252246886491776, + 0.060716334730386734, + 0.10059885680675507, + 0.14674191176891327 + ], + "credit_map": [ + 0.007607345934957266, + 0.014186087995767593, + 0.011288419365882874, + 0.008647091686725616, + 0.01847884990274906, + 0.007579934783279896, + 0.01098190899938345, + 0.007409702520817518, + 0.004382547922432423, + 0.0049404725432395935, + 0.004939926788210869, + 0.007066233549267054, + 0.007374766748398542, + 0.006041147280484438, + 0.008847116492688656, + 0.007533375173807144, + 0.01346422266215086, + 0.028442425653338432, + 0.016269978135824203, + 0.11257712543010712, + 0.018414417281746864, + 0.017840106040239334, + 0.015776682645082474, + 0.07626505941152573, + 0.006651341449469328, + 0.011285942047834396, + 0.00435038935393095, + 0.007366634439677, + 0.007495956029742956, + 0.004471933003515005, + 0.005352768115699291, + 0.005325707141309977, + 0.0135126868262887, + 0.013781309127807617, + 0.009872585535049438, + 0.010229111649096012, + 0.017859725281596184, + 0.012735394760966301, + 0.008991377428174019, + 0.0077628702856600285, + 0.008319389075040817, + 0.005348710808902979, + 0.007366825360804796, + 0.006286467891186476, + 0.005628429353237152, + 0.00504246586933732, + 0.006088273599743843, + 0.007209442090243101, + 0.008066846989095211, + 0.00530630350112915, + 0.009143003262579441, + 0.00961872935295105, + 0.007165163289755583, + 0.00738882040604949, + 0.00518210232257843, + 0.0053812782280147076, + 0.007380037568509579, + 0.005124500021338463, + 0.00899049174040556, + 0.006273801904171705, + 0.01050629559904337, + 0.010260318405926228, + 0.005916193593293428, + 0.006264691706746817, + 0.01401660405099392, + 0.006521044764667749, + 0.015192434191703796, + 0.010363857261836529, + 0.008828798308968544, + 0.009944348596036434, + 0.008605887182056904, + 0.027125881984829903, + 0.015601134859025478, + 0.012856549583375454, + 0.017334645614027977, + 0.015795258805155754, + 0.0314287394285202, + 0.014614504761993885, + 0.025694049894809723, + 0.013417033478617668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.029839513823390007 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189646303653717 + ], + "opposite_logprob_tokens": [ + -4.018189430236816 + ], + "logprob_reward": -0.018189646303653717, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018189646303653717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30207598209381104, + 0.03632485866546631, + 0.01747867465019226, + 0.06004142016172409, + 0.07516385614871979, + 0.0687662735581398, + 0.049530066549777985, + 0.0788498967885971, + 0.09103672206401825, + 0.22073222696781158 + ], + "credit_map": [ + 0.035081591457128525, + 0.0398038774728775, + 0.05213659629225731, + 0.034884262830019, + 0.03765672445297241, + 0.02329283580183983, + 0.0495433434844017, + 0.02967672608792782, + 0.00289027183316648, + 0.006741036660969257, + 0.003559228964149952, + 0.004911408293992281, + 0.003378482535481453, + 0.004966940730810165, + 0.0036837521474808455, + 0.006193738896399736, + 0.002169250976294279, + 0.002178907860070467, + 0.004128181841224432, + 0.0016617679502815008, + 0.0012807128950953484, + 0.0017197559354826808, + 0.001585585530847311, + 0.0027545106131583452, + 0.006156943738460541, + 0.0099054966121912, + 0.011454030871391296, + 0.009109887294471264, + 0.005427352152764797, + 0.005059301387518644, + 0.006035966333001852, + 0.006892440374940634, + 0.009305602870881557, + 0.006564796436578035, + 0.00950336828827858, + 0.008203503675758839, + 0.007241264916956425, + 0.009263062849640846, + 0.013912011869251728, + 0.011170245707035065, + 0.007620818447321653, + 0.006155282724648714, + 0.007979612797498703, + 0.02151617407798767, + 0.005498609971255064, + 0.003928209654986858, + 0.006340495310723782, + 0.009727069176733494, + 0.0077356453984975815, + 0.004334728233516216, + 0.01017673872411251, + 0.006271593738347292, + 0.006401672028005123, + 0.006813412997871637, + 0.0031203257385641336, + 0.004675951786339283, + 0.009721849113702774, + 0.006883169524371624, + 0.013035446405410767, + 0.009036616422235966, + 0.01247879583388567, + 0.011904067359864712, + 0.007568432483822107, + 0.008221525698900223, + 0.01129627414047718, + 0.006248092744499445, + 0.011339918710291386, + 0.009857764467597008, + 0.010758637450635433, + 0.010129516944289207, + 0.007458715699613094, + 0.023947807028889656, + 0.02323787473142147, + 0.01854569837450981, + 0.026301756501197815, + 0.024482691660523415, + 0.046956535428762436, + 0.02328612096607685, + 0.037790123373270035, + 0.020131435245275497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.018189646303653717 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14286410808563232 + ], + "opposite_logprob_tokens": [ + -2.017864227294922 + ], + "logprob_reward": -0.14286410808563232, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14286410808563232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03602931275963783, + 0.014972619712352753, + 0.13393646478652954, + 0.03796498477458954, + 0.10398237407207489, + 0.05882837623357773, + 0.04595312848687172, + 0.1300649493932724, + 0.09287264198064804, + 0.34539511799812317 + ], + "credit_map": [ + 0.0032566252630203962, + 0.00659950589761138, + 0.004978830926120281, + 0.003887275466695428, + 0.006188930012285709, + 0.0030221263878047466, + 0.005199138540774584, + 0.0028968832921236753, + 0.0012057202402502298, + 0.002053888514637947, + 0.0017673135735094547, + 0.0024476952385157347, + 0.002020919229835272, + 0.0017327385721728206, + 0.0021180452313274145, + 0.0016263012075796723, + 0.00729366485029459, + 0.00988729763776064, + 0.01663832552731037, + 0.03761918470263481, + 0.007328315172344446, + 0.00876396894454956, + 0.0054803285747766495, + 0.040925394743680954, + 0.0024845029693096876, + 0.01329272985458374, + 0.00257434812374413, + 0.007623421028256416, + 0.003496158868074417, + 0.003538998542353511, + 0.0020631756633520126, + 0.002891652984544635, + 0.01221533864736557, + 0.013336827978491783, + 0.009815847501158714, + 0.011133563704788685, + 0.02006429247558117, + 0.01703561656177044, + 0.01185024343430996, + 0.008530664257705212, + 0.006359124090522528, + 0.004527646582573652, + 0.005422776564955711, + 0.015975823625922203, + 0.006036337465047836, + 0.0046681794337928295, + 0.0052356719970703125, + 0.010602820664644241, + 0.0060838391073048115, + 0.004520084243267775, + 0.006900010164827108, + 0.006337897386401892, + 0.007489154115319252, + 0.0035558745730668306, + 0.005547119304537773, + 0.005519155878573656, + 0.016394583508372307, + 0.011682163923978806, + 0.016527974978089333, + 0.012179155834019184, + 0.02382776141166687, + 0.024551307782530785, + 0.012137667275965214, + 0.012764343060553074, + 0.010441088117659092, + 0.007453477941453457, + 0.010095963254570961, + 0.009438898414373398, + 0.014098888263106346, + 0.010517682880163193, + 0.007671419996768236, + 0.023155231028795242, + 0.035850051790475845, + 0.027494538575410843, + 0.04195694997906685, + 0.03592289984226227, + 0.07458319514989853, + 0.038254134356975555, + 0.05862317606806755, + 0.032710202038288116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.14286410808563232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079574935138226 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079574935138226, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079574935138226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059146761894226074, + 0.031826358288526535, + 0.08316559344530106, + 0.10909093171358109, + 0.24094423651695251, + 0.09297468513250351, + 0.05628300830721855, + 0.05796185880899429, + 0.15199342370033264, + 0.11661308258771896 + ], + "credit_map": [ + 0.004621703643351793, + 0.007179889362305403, + 0.007112778723239899, + 0.007308230735361576, + 0.008115618489682674, + 0.009445156902074814, + 0.010300841182470322, + 0.005062541924417019, + 0.0025285854935646057, + 0.003305074293166399, + 0.003005929058417678, + 0.00492573669180274, + 0.00458903843536973, + 0.00432269461452961, + 0.0052222516387701035, + 0.0039270478300750256, + 0.005097921937704086, + 0.007509161718189716, + 0.012032663449645042, + 0.032549675554037094, + 0.0038144218269735575, + 0.004196958150714636, + 0.0043813264928758144, + 0.013583463616669178, + 0.0068932813592255116, + 0.039232589304447174, + 0.009844549000263214, + 0.02007330022752285, + 0.01098706666380167, + 0.008773518726229668, + 0.0052979858592152596, + 0.007988648489117622, + 0.025049133226275444, + 0.029386006295681, + 0.022085903212428093, + 0.027832144871354103, + 0.043696336448192596, + 0.03835081681609154, + 0.031901560723781586, + 0.02264232374727726, + 0.012999550439417362, + 0.009872399270534515, + 0.010252883657813072, + 0.012404839508235455, + 0.009368747472763062, + 0.00961637869477272, + 0.008235638029873371, + 0.02022423781454563, + 0.007021013647317886, + 0.0051693967543542385, + 0.007613386493176222, + 0.01062651164829731, + 0.006279509514570236, + 0.00599302351474762, + 0.006714399438351393, + 0.006865768227726221, + 0.0068527162075042725, + 0.005906915757805109, + 0.0050154090858995914, + 0.004558316431939602, + 0.011078550480306149, + 0.0117623470723629, + 0.005991397425532341, + 0.006796208675950766, + 0.017714522778987885, + 0.010390611365437508, + 0.016973411664366722, + 0.016671648249030113, + 0.01912589929997921, + 0.017350681126117706, + 0.015552605502307415, + 0.0382140576839447, + 0.011942824348807335, + 0.01108669675886631, + 0.012982861138880253, + 0.011175528168678284, + 0.025568068027496338, + 0.012094441801309586, + 0.02095045894384384, + 0.010812207125127316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.011079574935138226 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tightly packed portrait-style close-up centers on the chest area of an unseen figure, where ethereal streams of fiery purple energy rapidly gather and weave into a shimmering, translucent bubble of magical energy—the Banshee's Veil. Its surface, a delicate crystalline grid, hums with contained power, casting a soft purple glow on the invisible armor beneath. Just as the sphere reaches its full, perfect form, a burst of searing crimson energy radiates from off-screen, violently colliding with the barrier and triggering a breathtaking chain reaction of deep purple light that ripples across the shield's surface in concentric waves, causing the entire structure to buckle and twist under the immense force. As the shield steadily and deliberately dissipates, fragmenting into a series of shimmering, fading points of light, the camera smoothly tilts upward, rising above the ornate dark steel armor to reveal the resolute face of a female warrior. Her face, framed by wisps of sweat-stained, raven-black hair, freed from its tight braid, wore a mask of concentration; her sharp, high cheekbones were illuminated by the residual purple light, her jaw set, her sharp grey eyes narrowed, unblinking, staring resolutely ahead, creating a tense and heroic atmosphere of a narrow victory.", + "seed": 12276, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_11/index_3592_seed_12276_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05181329846382141, + "vlm_logprob": -0.05181329846382141, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.05181329846382141, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1688830405473709 + ], + "opposite_logprob_tokens": [ + -1.9188830852508545 + ], + "logprob_reward": -0.1688830405473709, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1688830405473709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09047584980726242, + 0.07885055989027023, + 0.020141247659921646, + 0.4171586334705353, + 0.10053602606058121, + 0.0554351881146431, + 0.059659913182258606, + 0.05156787857413292, + 0.09714386612176895, + 0.02903090789914131 + ], + "credit_map": [ + 0.006743913516402245, + 0.010603079572319984, + 0.005928788334131241, + 0.02415706403553486, + 0.007591166067868471, + 0.005450309254229069, + 0.021439941599965096, + 0.008561581373214722, + 0.005497668869793415, + 0.006894673220813274, + 0.007583890110254288, + 0.006104916334152222, + 0.026391953229904175, + 0.005529820919036865, + 0.015232265926897526, + 0.005615371745079756, + 0.0015976576833054423, + 0.0025703164283186197, + 0.002495549852028489, + 0.003148735500872135, + 0.0027970443479716778, + 0.0015084664337337017, + 0.0022586421109735966, + 0.0037648347206413746, + 0.02109081856906414, + 0.02676200121641159, + 0.024195808917284012, + 0.0217296089977026, + 0.029033590108156204, + 0.016860565170645714, + 0.16043570637702942, + 0.11705050617456436, + 0.010649660602211952, + 0.013470993377268314, + 0.013899187557399273, + 0.015626946464180946, + 0.01508569996803999, + 0.009816588833928108, + 0.012399065308272839, + 0.009587886743247509, + 0.006702420301735401, + 0.006769613828510046, + 0.005137488711625338, + 0.011023231782019138, + 0.00493048969656229, + 0.005832223687320948, + 0.008175489492714405, + 0.006864228285849094, + 0.005534226540476084, + 0.00902518816292286, + 0.005969781894236803, + 0.005803728476166725, + 0.008953376673161983, + 0.008846377953886986, + 0.007223708089441061, + 0.008303525857627392, + 0.0038089037407189608, + 0.003467948641628027, + 0.0034480192698538303, + 0.0033068854827433825, + 0.005882616154849529, + 0.002487560035660863, + 0.016370495781302452, + 0.01279545109719038, + 0.009010293520987034, + 0.007672418840229511, + 0.009502417407929897, + 0.0072632464580237865, + 0.012406600639224052, + 0.010644149966537952, + 0.026974136009812355, + 0.01367060374468565, + 0.002630282426252961, + 0.004091593902558088, + 0.004070595372468233, + 0.0040862541645765305, + 0.0034025865606963634, + 0.003032980253919959, + 0.004181450232863426, + 0.003535163588821888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.1688830405473709 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026390315964818 + ], + "opposite_logprob_tokens": [ + -3.651390314102173 + ], + "logprob_reward": -0.026390315964818, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026390315964818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15702158212661743, + 0.1291128545999527, + 0.048592258244752884, + 0.11270014196634293, + 0.04410041123628616, + 0.12932319939136505, + 0.09196418523788452, + 0.06046651303768158, + 0.07585014402866364, + 0.1508687287569046 + ], + "credit_map": [ + 0.010704326443374157, + 0.011816958896815777, + 0.021476972848176956, + 0.01997329294681549, + 0.01814611628651619, + 0.040201444178819656, + 0.019033579155802727, + 0.01566888764500618, + 0.0058608404360711575, + 0.0075942836701869965, + 0.00908108800649643, + 0.006435072515159845, + 0.03621735796332359, + 0.012596555054187775, + 0.04179123789072037, + 0.009536403231322765, + 0.006049540359526873, + 0.005848516244441271, + 0.00596390338614583, + 0.006027628667652607, + 0.007879829034209251, + 0.004778462927788496, + 0.005732869729399681, + 0.006311508361250162, + 0.020604213699698448, + 0.013426188379526138, + 0.011765778064727783, + 0.010199097916483879, + 0.017331218346953392, + 0.009498034603893757, + 0.013131022453308105, + 0.01674458011984825, + 0.004450061358511448, + 0.005841571372002363, + 0.0043764919973909855, + 0.005645801778882742, + 0.0041870009154081345, + 0.006138720083981752, + 0.0063352216966450214, + 0.007125540636479855, + 0.009998100809752941, + 0.010326010175049305, + 0.028415998443961143, + 0.016461774706840515, + 0.013118580915033817, + 0.010415351949632168, + 0.031986214220523834, + 0.008601157926023006, + 0.005325892474502325, + 0.019316891208291054, + 0.007906332612037659, + 0.014606645330786705, + 0.011439159512519836, + 0.007177120074629784, + 0.01425633393228054, + 0.011935804039239883, + 0.006951962597668171, + 0.006930157542228699, + 0.006814878899604082, + 0.008569246158003807, + 0.008288943208754063, + 0.008065654896199703, + 0.008525841869413853, + 0.006319819949567318, + 0.013026528060436249, + 0.005297020077705383, + 0.004374272655695677, + 0.004809828475117683, + 0.00758123304694891, + 0.013999590650200844, + 0.01573576033115387, + 0.011025902815163136, + 0.02350202575325966, + 0.014070381410419941, + 0.012303316034376621, + 0.016446970403194427, + 0.01835002563893795, + 0.03664607182145119, + 0.015435315668582916, + 0.01411460805684328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.026390315964818 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082051321864128 + ], + "opposite_logprob_tokens": [ + -4.511082172393799 + ], + "logprob_reward": -0.011082051321864128, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082051321864128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04817001894116402, + 0.0894559994339943, + 0.02799864113330841, + 0.20069439709186554, + 0.0445706732571125, + 0.07512149959802628, + 0.23643551766872406, + 0.06701965630054474, + 0.056749649345874786, + 0.15378396213054657 + ], + "credit_map": [ + 0.005259704310446978, + 0.005387639626860619, + 0.00426728930324316, + 0.012091871351003647, + 0.004633319564163685, + 0.005289937369525433, + 0.0061967093497514725, + 0.005043548997491598, + 0.004557900596410036, + 0.005427032709121704, + 0.005963824689388275, + 0.005695017520338297, + 0.025445884093642235, + 0.006552514154464006, + 0.030161185190081596, + 0.005652647465467453, + 0.0026975846849381924, + 0.0035575288347899914, + 0.003315672976896167, + 0.0031679205130785704, + 0.006625316571444273, + 0.0026245517656207085, + 0.0027403163257986307, + 0.003269749926403165, + 0.02217765711247921, + 0.014940469525754452, + 0.015746962279081345, + 0.009788135066628456, + 0.018804563209414482, + 0.010620689950883389, + 0.06761150807142258, + 0.04100441187620163, + 0.0042709349654614925, + 0.005130804609507322, + 0.004015019163489342, + 0.005405989475548267, + 0.005070252809673548, + 0.005847719497978687, + 0.007184086367487907, + 0.007645870093256235, + 0.008091608993709087, + 0.0117123331874609, + 0.008107359521090984, + 0.010315528139472008, + 0.007081982214003801, + 0.009134863503277302, + 0.013160469010472298, + 0.007517355028539896, + 0.020229270681738853, + 0.045111846178770065, + 0.028430547565221786, + 0.04373186454176903, + 0.027666525915265083, + 0.018910855054855347, + 0.028681837022304535, + 0.02367275394499302, + 0.006867669522762299, + 0.008677802979946136, + 0.005800908897072077, + 0.0072331842966377735, + 0.010132004506886005, + 0.007966538891196251, + 0.011046914383769035, + 0.009294634684920311, + 0.004660436883568764, + 0.007057988550513983, + 0.007715252693742514, + 0.008591665886342525, + 0.0071905218064785, + 0.00458678649738431, + 0.011263836175203323, + 0.005683162249624729, + 0.025200646370649338, + 0.011731712147593498, + 0.01481997687369585, + 0.014925611205399036, + 0.015399067662656307, + 0.03864780440926552, + 0.016721786931157112, + 0.016337351873517036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.011082051321864128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026365818455815315 + ], + "opposite_logprob_tokens": [ + -3.6513657569885254 + ], + "logprob_reward": -0.026365818455815315, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026365818455815315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03285084292292595, + 0.014582398347556591, + 0.010524901561439037, + 0.12462808191776276, + 0.019946804270148277, + 0.04665528982877731, + 0.1597081571817398, + 0.06497329473495483, + 0.09155649691820145, + 0.4345737099647522 + ], + "credit_map": [ + 0.0024267747066915035, + 0.003858884098008275, + 0.0024295358452945948, + 0.0077508362010121346, + 0.0034063104540109634, + 0.005363646429032087, + 0.004052779637277126, + 0.0035620795097202063, + 0.0013055785093456507, + 0.001558526768349111, + 0.001281743636354804, + 0.0015765337739139795, + 0.0029386081732809544, + 0.0013388285879045725, + 0.0023109253961592913, + 0.0022716522216796875, + 0.001118925865739584, + 0.0011885040439665318, + 0.0015790386823937297, + 0.0011838431237265468, + 0.002511254046112299, + 0.000883374537806958, + 0.0009323748527094722, + 0.0011275869328528643, + 0.0050080688670277596, + 0.004720616620033979, + 0.005598742980509996, + 0.0047804513014853, + 0.006584621034562588, + 0.006480022799223661, + 0.0482085682451725, + 0.04324699938297272, + 0.0015380809782072902, + 0.0016006254591047764, + 0.0017982734134420753, + 0.002376399002969265, + 0.0016623022966086864, + 0.0023616086691617966, + 0.0037250297609716654, + 0.004884485155344009, + 0.006252515595406294, + 0.006717681419104338, + 0.00513213686645031, + 0.006492712534964085, + 0.005531400442123413, + 0.004989904817193747, + 0.006679800804704428, + 0.004859135486185551, + 0.016442997381091118, + 0.020009486004710197, + 0.016812996938824654, + 0.021458851173520088, + 0.02716589719057083, + 0.020567094907164574, + 0.01960473321378231, + 0.01764610782265663, + 0.004217144101858139, + 0.007541941478848457, + 0.0041922577656805515, + 0.005146597977727652, + 0.009134217165410519, + 0.004066520370543003, + 0.017527475953102112, + 0.013147147372364998, + 0.007852497510612011, + 0.014260956086218357, + 0.007296511437743902, + 0.00987845566123724, + 0.0158319640904665, + 0.014182434417307377, + 0.012982026673853397, + 0.009271645918488503, + 0.06229550391435623, + 0.04988479241728783, + 0.03611689433455467, + 0.04903678968548775, + 0.05365193262696266, + 0.09916844218969345, + 0.05630650743842125, + 0.028112905099987984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.026365818455815315 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263452660292387 + ], + "opposite_logprob_tokens": [ + -3.6513452529907227 + ], + "logprob_reward": -0.0263452660292387, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263452660292387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13372740149497986, + 0.02410465106368065, + 0.030432170256972313, + 0.12500768899917603, + 0.023874087259173393, + 0.042927395552396774, + 0.10879746824502945, + 0.08916263282299042, + 0.2307448834180832, + 0.19122159481048584 + ], + "credit_map": [ + 0.006984620355069637, + 0.01584175042808056, + 0.009218069724738598, + 0.04029588773846626, + 0.013269376941025257, + 0.02237829752266407, + 0.017632655799388885, + 0.008106771856546402, + 0.002343114698305726, + 0.0021628455724567175, + 0.0032378463074564934, + 0.0026773489080369473, + 0.003245197469368577, + 0.0026757053565233946, + 0.0032544045243412256, + 0.004508189857006073, + 0.003586366306990385, + 0.0038988180458545685, + 0.00319349835626781, + 0.003548482432961464, + 0.007282419130206108, + 0.003077675821259618, + 0.002484579337760806, + 0.003360332455486059, + 0.014015230350196362, + 0.01194316241890192, + 0.009803160093724728, + 0.007077357266098261, + 0.009914344176650047, + 0.0073473211377859116, + 0.03766946494579315, + 0.027237668633461, + 0.001691228593699634, + 0.0024647777900099754, + 0.002549239667132497, + 0.003601343836635351, + 0.002116501098498702, + 0.002632108284160495, + 0.0044633131474256516, + 0.004355576355010271, + 0.0035778810270130634, + 0.005986880045384169, + 0.0039877560921013355, + 0.007681403309106827, + 0.004039119463413954, + 0.003933845553547144, + 0.008684792555868626, + 0.005035725422203541, + 0.00795214157551527, + 0.009726081974804401, + 0.007906213402748108, + 0.03749433159828186, + 0.008607807569205761, + 0.007564846891909838, + 0.019563574343919754, + 0.009982484392821789, + 0.006557846441864967, + 0.014033912681043148, + 0.01247908640652895, + 0.012610354460775852, + 0.01341169886291027, + 0.00801294669508934, + 0.012657134793698788, + 0.009399659931659698, + 0.0316087082028389, + 0.018823722377419472, + 0.015084720216691494, + 0.016255903989076614, + 0.03646448627114296, + 0.020407436415553093, + 0.06825322657823563, + 0.023846706375479698, + 0.019410153850913048, + 0.027532516047358513, + 0.020859340205788612, + 0.02854418009519577, + 0.02077948860824108, + 0.0335855670273304, + 0.023312710225582123, + 0.01719764806330204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0263452660292387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tightly packed portrait-style close-up centers on the chest area of an unseen figure, where ethereal streams of fiery purple energy rapidly gather and weave into a shimmering, translucent bubble of magical energy—the Banshee's Veil. Its surface, a delicate crystalline grid, hums with contained power, casting a soft purple glow on the invisible armor beneath. Just as the sphere reaches its full, perfect form, a burst of searing crimson energy radiates from off-screen, violently colliding with the barrier and triggering a breathtaking chain reaction of deep purple light that ripples across the shield's surface in concentric waves, causing the entire structure to buckle and twist under the immense force. As the shield steadily and deliberately dissipates, fragmenting into a series of shimmering, fading points of light, the camera smoothly tilts upward, rising above the ornate dark steel armor to reveal the resolute face of a female warrior. Her face, framed by wisps of sweat-stained, raven-black hair, freed from its tight braid, wore a mask of concentration; her sharp, high cheekbones were illuminated by the residual purple light, her jaw set, her sharp grey eyes narrowed, unblinking, staring resolutely ahead, creating a tense and heroic atmosphere of a narrow victory.", + "seed": 12277, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_11/index_3592_seed_12277_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025522243604063986, + "vlm_logprob": -0.025522243604063986, + "vlm_yes_no_margin": 4.0749999523162845, + "vlm_token_credit": -0.025522243604063986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006951788440346718 + ], + "opposite_logprob_tokens": [ + -5.506951808929443 + ], + "logprob_reward": -0.006951788440346718, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006951788440346718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22429092228412628, + 0.1402306854724884, + 0.03473953157663345, + 0.1053772121667862, + 0.03213464841246605, + 0.03860597684979439, + 0.05576257035136223, + 0.04225287213921547, + 0.0642591118812561, + 0.26234638690948486 + ], + "credit_map": [ + 0.018554475158452988, + 0.020528409630060196, + 0.027399176731705666, + 0.022420011460781097, + 0.03342408314347267, + 0.03208421170711517, + 0.02300380729138851, + 0.04687676578760147, + 0.012944260612130165, + 0.011358908377587795, + 0.022172749042510986, + 0.04449254274368286, + 0.017231954261660576, + 0.009385780431330204, + 0.00904577411711216, + 0.013598711229860783, + 0.003663190407678485, + 0.004837182816118002, + 0.0033720177598297596, + 0.0038082068786025047, + 0.004060965497046709, + 0.0047563291154801846, + 0.0065951100550591946, + 0.003646531840786338, + 0.006731178145855665, + 0.012630995362997055, + 0.003731738543137908, + 0.004994513001292944, + 0.005633953493088484, + 0.03815300762653351, + 0.005238259211182594, + 0.028263572603464127, + 0.0025319133419543505, + 0.0054854173213243484, + 0.0036162277683615685, + 0.004934187512844801, + 0.003958006855100393, + 0.0027079381980001926, + 0.004314360674470663, + 0.004586595110595226, + 0.004233429674059153, + 0.0054490710608661175, + 0.0032734910491853952, + 0.006087960675358772, + 0.004164283629506826, + 0.0038002864457666874, + 0.004578946623951197, + 0.0070185065269470215, + 0.006406635511666536, + 0.008561369962990284, + 0.005786198191344738, + 0.00816053245216608, + 0.008463403210043907, + 0.005530218128114939, + 0.00804171059280634, + 0.004812503699213266, + 0.005839396733790636, + 0.0049787224270403385, + 0.004671825561672449, + 0.003686661832034588, + 0.007871433161199093, + 0.005175442900508642, + 0.006207821890711784, + 0.003821565303951502, + 0.006196668837219477, + 0.00917569175362587, + 0.006666809320449829, + 0.005713213700801134, + 0.0144572788849473, + 0.007776624523103237, + 0.00836670957505703, + 0.005906115286052227, + 0.023051291704177856, + 0.008027534000575542, + 0.046811144798994064, + 0.010315501131117344, + 0.04896514490246773, + 0.017537958920001984, + 0.09608783572912216, + 0.011549958027899265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.006951788440346718 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014225034043192863 + ], + "opposite_logprob_tokens": [ + -4.264225006103516 + ], + "logprob_reward": -0.014225034043192863, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014225034043192863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09782381355762482, + 0.09162794798612595, + 0.07499004900455475, + 0.13819707930088043, + 0.04365343973040581, + 0.15018929541110992, + 0.05308489501476288, + 0.07161348313093185, + 0.2172377109527588, + 0.061582278460264206 + ], + "credit_map": [ + 0.010144054889678955, + 0.011510995216667652, + 0.01049541775137186, + 0.012066084891557693, + 0.008432306349277496, + 0.015349850989878178, + 0.014590979553759098, + 0.015234125778079033, + 0.00800917949527502, + 0.011710289865732193, + 0.011089303530752659, + 0.015870695933699608, + 0.01409170962870121, + 0.012081976979970932, + 0.008376560173928738, + 0.010398237966001034, + 0.00515437126159668, + 0.01452227309346199, + 0.005817420314997435, + 0.007837433367967606, + 0.007745034527033567, + 0.01803976111114025, + 0.00949067436158657, + 0.006383081898093224, + 0.005960382055491209, + 0.048454806208610535, + 0.005332091357558966, + 0.0053905039094388485, + 0.009806249290704727, + 0.04429545998573303, + 0.008512006141245365, + 0.010445586405694485, + 0.0032294418197125196, + 0.007692410610616207, + 0.006045233923941851, + 0.006238562520593405, + 0.005733979865908623, + 0.0031712541822344065, + 0.00646914541721344, + 0.005073410924524069, + 0.013106455095112324, + 0.014944061636924744, + 0.01337987557053566, + 0.019285814836621284, + 0.020625246688723564, + 0.013538115657866001, + 0.015472542494535446, + 0.039837174117565155, + 0.00671680923551321, + 0.006342611275613308, + 0.005099032539874315, + 0.004795251879841089, + 0.012707578018307686, + 0.004679819568991661, + 0.008232000283896923, + 0.004511792678385973, + 0.008255151100456715, + 0.008476164191961288, + 0.008738275617361069, + 0.0073036085814237595, + 0.012171492911875248, + 0.009324882179498672, + 0.010449055582284927, + 0.006894852500408888, + 0.02148105949163437, + 0.027217362076044083, + 0.024471722543239594, + 0.018884768709540367, + 0.04456980153918266, + 0.02211298607289791, + 0.042388249188661575, + 0.016111763194203377, + 0.005822476930916309, + 0.007493985816836357, + 0.007021746598184109, + 0.006093711592257023, + 0.0100623844191432, + 0.005171294789761305, + 0.014964181929826736, + 0.004952500108629465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.014225034043192863 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819046586751938 + ], + "opposite_logprob_tokens": [ + -4.018190383911133 + ], + "logprob_reward": -0.01819046586751938, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819046586751938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07358816266059875, + 0.05114186182618141, + 0.03382544219493866, + 0.039477553218603134, + 0.0246651079505682, + 0.11568431556224823, + 0.09451740980148315, + 0.07310526072978973, + 0.24299000203609467, + 0.2510049045085907 + ], + "credit_map": [ + 0.008086372166872025, + 0.008389921858906746, + 0.009792580269277096, + 0.0068114507012069225, + 0.009636342525482178, + 0.01238227542489767, + 0.006278637330979109, + 0.012210581451654434, + 0.004920023959130049, + 0.005950957536697388, + 0.007759931962937117, + 0.012024858966469765, + 0.004894095938652754, + 0.0056242793798446655, + 0.004066006280481815, + 0.005901704076677561, + 0.0031003362964838743, + 0.005146115552634001, + 0.003001550678163767, + 0.004353949334472418, + 0.004559903405606747, + 0.005707481410354376, + 0.004559866618365049, + 0.0033962400630116463, + 0.0035215881653130054, + 0.010304238647222519, + 0.002576913218945265, + 0.003056830959394574, + 0.0031135091558098793, + 0.007134380284696817, + 0.0035894070751965046, + 0.006180687341839075, + 0.002526768483221531, + 0.00314139761030674, + 0.002690599299967289, + 0.003955232445150614, + 0.00430260319262743, + 0.0019182950491085649, + 0.0030130944214761257, + 0.003117117565125227, + 0.012066948227584362, + 0.0146914292126894, + 0.010146214626729488, + 0.012822313234210014, + 0.01325606182217598, + 0.013744276016950607, + 0.011224945075809956, + 0.02773212641477585, + 0.012081777676939964, + 0.013202336616814137, + 0.009240536019206047, + 0.009452407248318195, + 0.016867676749825478, + 0.009469155222177505, + 0.014669141732156277, + 0.009534373879432678, + 0.006119654513895512, + 0.00929765123873949, + 0.007230727467685938, + 0.007555351592600346, + 0.01797088421881199, + 0.006944520398974419, + 0.011619256809353828, + 0.006367219612002373, + 0.025124505162239075, + 0.037169475108385086, + 0.03036370873451233, + 0.019948827102780342, + 0.056806959211826324, + 0.024458138272166252, + 0.032473448663949966, + 0.016644928604364395, + 0.020921088755130768, + 0.012226478196680546, + 0.041800547391176224, + 0.010923507623374462, + 0.04731031507253647, + 0.016002127900719643, + 0.08995930105447769, + 0.011861540377140045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.01819046586751938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005627453327179 + ], + "opposite_logprob_tokens": [ + -2.695056200027466 + ], + "logprob_reward": -0.07005627453327179, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005627453327179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06360696256160736, + 0.019115624949336052, + 0.02114260010421276, + 0.032578613609075546, + 0.009803444147109985, + 0.019271105527877808, + 0.041734158992767334, + 0.1870994120836258, + 0.41357725858688354, + 0.1920708417892456 + ], + "credit_map": [ + 0.0068241748958826065, + 0.0049082813784480095, + 0.007821062579751015, + 0.0046237315982580185, + 0.009334830567240715, + 0.008915557526051998, + 0.008832757361233234, + 0.01234656386077404, + 0.002739069750532508, + 0.0017316282028332353, + 0.003013204550370574, + 0.004519484471529722, + 0.001334657077677548, + 0.0022153242025524378, + 0.0015720488736405969, + 0.001990208402276039, + 0.0015966999344527721, + 0.004249802324920893, + 0.0017131507629528642, + 0.002380103338509798, + 0.0018644449301064014, + 0.004718462470918894, + 0.00296253920532763, + 0.0016573966713622212, + 0.0015985232312232256, + 0.010694481432437897, + 0.0011470046592876315, + 0.001788519904948771, + 0.0018353257328271866, + 0.011639616452157497, + 0.00179822719655931, + 0.00207691197283566, + 0.001118967542424798, + 0.0014687278307974339, + 0.0014110964257270098, + 0.0013808381045237184, + 0.0013723088195547462, + 0.0008489997708238661, + 0.0012785540893673897, + 0.0009239513310603797, + 0.0020021304953843355, + 0.002783093135803938, + 0.0022080661728978157, + 0.002195141278207302, + 0.002062868094071746, + 0.0018347601871937513, + 0.001948406919836998, + 0.004236639477312565, + 0.004599238280206919, + 0.005773359443992376, + 0.005671789404004812, + 0.004787394776940346, + 0.006743881851434708, + 0.003827318549156189, + 0.006507343612611294, + 0.0038238344714045525, + 0.016683422029018402, + 0.026042401790618896, + 0.02043144218623638, + 0.018181435763835907, + 0.03441113233566284, + 0.022393502295017242, + 0.027786560356616974, + 0.021169522777199745, + 0.04068262130022049, + 0.05632846802473068, + 0.046144064515829086, + 0.03666695952415466, + 0.08642882108688354, + 0.04236869141459465, + 0.07400514930486679, + 0.030952490866184235, + 0.015670809894800186, + 0.0134940380230546, + 0.029339205473661423, + 0.01368195191025734, + 0.0333297960460186, + 0.015425066463649273, + 0.05621195584535599, + 0.014918019063770771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.07005627453327179 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818765513598919 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.01818765513598919, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818765513598919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06235430762171745, + 0.03484378755092621, + 0.061835963279008865, + 0.06010466068983078, + 0.025071775540709496, + 0.03478559851646423, + 0.09970256686210632, + 0.05770750343799591, + 0.1329524964094162, + 0.4306413531303406 + ], + "credit_map": [ + 0.006768897175788879, + 0.006196032278239727, + 0.006751232780516148, + 0.009182651527225971, + 0.007387788034975529, + 0.006379078608006239, + 0.009777029044926167, + 0.009911599569022655, + 0.0031681053806096315, + 0.0037251508329063654, + 0.005883015226572752, + 0.008116481825709343, + 0.0029337676241993904, + 0.002793307416141033, + 0.004100319463759661, + 0.0041236416436731815, + 0.0033939285203814507, + 0.015309464186429977, + 0.003620087867602706, + 0.004411682020872831, + 0.005449544172734022, + 0.017515286803245544, + 0.007272160612046719, + 0.004863809794187546, + 0.00267992471344769, + 0.020790714770555496, + 0.0023116376250982285, + 0.003853193949908018, + 0.0039664143696427345, + 0.0166737399995327, + 0.003674281993880868, + 0.006154756061732769, + 0.0028104984667152166, + 0.002738100942224264, + 0.004202739801257849, + 0.0030794544145464897, + 0.003755951765924692, + 0.002107560634613037, + 0.0032477257773280144, + 0.0031297430396080017, + 0.005621364340186119, + 0.005680297501385212, + 0.0036333734169602394, + 0.005085830111056566, + 0.002817694563418627, + 0.003595584537833929, + 0.004429060500115156, + 0.00392239261418581, + 0.009829340502619743, + 0.014985761605203152, + 0.007607211824506521, + 0.012622670270502567, + 0.017142854630947113, + 0.011427639052271843, + 0.015454812906682491, + 0.010632270015776157, + 0.005569347646087408, + 0.006837925873696804, + 0.007616217713803053, + 0.005533360876142979, + 0.010078297927975655, + 0.007351519074290991, + 0.007717622444033623, + 0.007003212813287973, + 0.013334308750927448, + 0.01938672550022602, + 0.014447434805333614, + 0.010501853190362453, + 0.0332280769944191, + 0.013348178938031197, + 0.01864452287554741, + 0.010061387903988361, + 0.0386362299323082, + 0.014958303421735764, + 0.0778779461979866, + 0.019014418125152588, + 0.0799010768532753, + 0.027347367256879807, + 0.1522456854581833, + 0.020660335198044777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.01818765513598919 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tightly packed portrait-style close-up centers on the chest area of an unseen figure, where ethereal streams of fiery purple energy rapidly gather and weave into a shimmering, translucent bubble of magical energy—the Banshee's Veil. Its surface, a delicate crystalline grid, hums with contained power, casting a soft purple glow on the invisible armor beneath. Just as the sphere reaches its full, perfect form, a burst of searing crimson energy radiates from off-screen, violently colliding with the barrier and triggering a breathtaking chain reaction of deep purple light that ripples across the shield's surface in concentric waves, causing the entire structure to buckle and twist under the immense force. As the shield steadily and deliberately dissipates, fragmenting into a series of shimmering, fading points of light, the camera smoothly tilts upward, rising above the ornate dark steel armor to reveal the resolute face of a female warrior. Her face, framed by wisps of sweat-stained, raven-black hair, freed from its tight braid, wore a mask of concentration; her sharp, high cheekbones were illuminated by the residual purple light, her jaw set, her sharp grey eyes narrowed, unblinking, staring resolutely ahead, creating a tense and heroic atmosphere of a narrow victory.", + "seed": 12278, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_11/index_3592_seed_12278_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.032664966955780986, + "vlm_logprob": -0.032664966955780986, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.032664966955780986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009387754835188389 + ], + "opposite_logprob_tokens": [ + -5.259387969970703 + ], + "logprob_reward": -0.009387754835188389, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.009387754835188389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22629882395267487, + 0.1993149071931839, + 0.22294966876506805, + 0.038037557154893875, + 0.0617673397064209, + 0.06841462850570679, + 0.042863599956035614, + 0.03237541764974594, + 0.06994586437940598, + 0.0380321703851223 + ], + "credit_map": [ + 0.01487648207694292, + 0.025129789486527443, + 0.01313840039074421, + 0.02117064595222473, + 0.04508773609995842, + 0.020287498831748962, + 0.07508230209350586, + 0.011525968089699745, + 0.007855817675590515, + 0.07028757035732269, + 0.005428431089967489, + 0.008837695233523846, + 0.01172198262065649, + 0.005782955791801214, + 0.07839218527078629, + 0.011008274741470814, + 0.00946676917374134, + 0.006308890413492918, + 0.00591887766495347, + 0.018431704491376877, + 0.010750805027782917, + 0.07692615687847137, + 0.08335819095373154, + 0.01178828440606594, + 0.0038359384052455425, + 0.003791940165683627, + 0.003546160412952304, + 0.004842980299144983, + 0.005267282482236624, + 0.004196778405457735, + 0.0038006494287401438, + 0.008755830116569996, + 0.009268369525671005, + 0.012882276438176632, + 0.004572063684463501, + 0.0057973964139819145, + 0.006128308828920126, + 0.004474520217627287, + 0.005446660332381725, + 0.01319774892181158, + 0.005688080098479986, + 0.01767781376838684, + 0.005569113418459892, + 0.015507852658629417, + 0.006531338673084974, + 0.004492839332669973, + 0.008505934849381447, + 0.004441660363227129, + 0.0071291569620370865, + 0.007048602681607008, + 0.00519903190433979, + 0.0067624496296048164, + 0.004437028896063566, + 0.003108271397650242, + 0.0038788809906691313, + 0.005300180520862341, + 0.002868558978661895, + 0.007805647794157267, + 0.002570049837231636, + 0.005995019339025021, + 0.00483290571719408, + 0.0027247266843914986, + 0.003379320725798607, + 0.0021991918329149485, + 0.014811606146395206, + 0.010704616084694862, + 0.006015915889292955, + 0.020275475457310677, + 0.003994643688201904, + 0.004574711434543133, + 0.004710833076387644, + 0.004858069587498903, + 0.003950860351324081, + 0.003965602722018957, + 0.00496643828228116, + 0.004468604922294617, + 0.004631297662854195, + 0.00508518423885107, + 0.004324057605117559, + 0.006640125997364521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.009387754835188389 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331940270960331 + ], + "opposite_logprob_tokens": [ + -3.7733194828033447 + ], + "logprob_reward": -0.02331940270960331, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331940270960331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07499732822179794, + 0.17707142233848572, + 0.3880355656147003, + 0.02196601592004299, + 0.0191714558750391, + 0.025775762274861336, + 0.1284644454717636, + 0.049204062670469284, + 0.07178889214992523, + 0.043525032699108124 + ], + "credit_map": [ + 0.0038236796390265226, + 0.006089446600526571, + 0.0030487393960356712, + 0.003717775223776698, + 0.022909436374902725, + 0.004541310016065836, + 0.026276426389813423, + 0.004590511787682772, + 0.008043687790632248, + 0.06238901987671852, + 0.005664549767971039, + 0.009316051378846169, + 0.01196730975061655, + 0.005775406025350094, + 0.06508062779903412, + 0.008834767155349255, + 0.014306118711829185, + 0.008516929112374783, + 0.007595621980726719, + 0.03392450511455536, + 0.019648374989628792, + 0.13375656306743622, + 0.15142029523849487, + 0.018867196515202522, + 0.0018441630527377129, + 0.003032105276361108, + 0.0018661182839423418, + 0.0035805553197860718, + 0.003625051584094763, + 0.003426275448873639, + 0.002362437080591917, + 0.0022293096408247948, + 0.0019475274020805955, + 0.0022494071163237095, + 0.002033904194831848, + 0.0030475291423499584, + 0.003438086248934269, + 0.0018785942811518908, + 0.0027860624250024557, + 0.0017903466941788793, + 0.0021686945110559464, + 0.003425880568102002, + 0.0030429468024522066, + 0.0036154408007860184, + 0.002036293502897024, + 0.00562556367367506, + 0.003453959012404084, + 0.0024069822393357754, + 0.025271089747548103, + 0.016697188839316368, + 0.016314078122377396, + 0.016489721834659576, + 0.01059557031840086, + 0.016658660024404526, + 0.014602686278522015, + 0.011835441924631596, + 0.0049563078209757805, + 0.009687202982604504, + 0.003985087852925062, + 0.008630658499896526, + 0.005762167740613222, + 0.004661640617996454, + 0.00633646035566926, + 0.005184536799788475, + 0.013817798346281052, + 0.011739635840058327, + 0.005626074969768524, + 0.022525891661643982, + 0.0048590656369924545, + 0.004808429162949324, + 0.00428242702037096, + 0.004129564389586449, + 0.004680098965764046, + 0.004757336340844631, + 0.0052585843950510025, + 0.006659470032900572, + 0.005938728339970112, + 0.005981402471661568, + 0.004726839251816273, + 0.00552257290109992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.02331940270960331 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197415672242641 + ], + "opposite_logprob_tokens": [ + -4.26419734954834 + ], + "logprob_reward": -0.014197415672242641, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197415672242641, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12371256947517395, + 0.299513041973114, + 0.1568288803100586, + 0.0287652425467968, + 0.03688568249344826, + 0.053924400359392166, + 0.09334869682788849, + 0.08828871697187424, + 0.05840858072042465, + 0.06032421067357063 + ], + "credit_map": [ + 0.006699718069285154, + 0.008380596525967121, + 0.005504746455699205, + 0.00668361596763134, + 0.03730953484773636, + 0.007960719987750053, + 0.04506230354309082, + 0.006111329421401024, + 0.012999257072806358, + 0.10498851537704468, + 0.008950838819146156, + 0.014526426792144775, + 0.020781317725777626, + 0.008586491458117962, + 0.11479263007640839, + 0.013887578621506691, + 0.006126854103058577, + 0.0033921110443770885, + 0.003860516706481576, + 0.013654964044690132, + 0.007895960472524166, + 0.05362861603498459, + 0.06054554507136345, + 0.007724305149167776, + 0.003244853112846613, + 0.003585123224183917, + 0.002331433817744255, + 0.005110980477184057, + 0.004576642531901598, + 0.0038165482692420483, + 0.003358502173796296, + 0.002741158241406083, + 0.004833504557609558, + 0.006745697930455208, + 0.0029301864560693502, + 0.0038336385041475296, + 0.005520742852240801, + 0.002992302179336548, + 0.004220319911837578, + 0.005809290800243616, + 0.00426869560033083, + 0.008387774229049683, + 0.005332121159881353, + 0.014083759859204292, + 0.00449095806106925, + 0.0055958605371415615, + 0.008750557899475098, + 0.0030146718490868807, + 0.016824061051011086, + 0.012250365689396858, + 0.01083111297339201, + 0.012712963856756687, + 0.007918175309896469, + 0.010860150679945946, + 0.013236060738563538, + 0.008715803734958172, + 0.00985329132527113, + 0.01591920480132103, + 0.010628453455865383, + 0.014963057823479176, + 0.009626355022192001, + 0.008765066042542458, + 0.008741146884858608, + 0.009792137891054153, + 0.006251310929656029, + 0.008973317220807076, + 0.0049637979827821255, + 0.006559673696756363, + 0.011125393211841583, + 0.006017220672219992, + 0.0075402152724564075, + 0.006977656856179237, + 0.006823574658483267, + 0.007502419408410788, + 0.00570263434201479, + 0.0074357884004712105, + 0.008724725805222988, + 0.008305595256388187, + 0.007218600250780582, + 0.00861087441444397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.014197415672242641 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035248100757599 + ], + "opposite_logprob_tokens": [ + -2.3503525257110596 + ], + "logprob_reward": -0.10035248100757599, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035248100757599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06099371984601021, + 0.10755511373281479, + 0.05300295725464821, + 0.013674605637788773, + 0.016142891719937325, + 0.10985718667507172, + 0.09047919511795044, + 0.14532843232154846, + 0.24551649391651154, + 0.157449409365654 + ], + "credit_map": [ + 0.003080765949562192, + 0.004317056387662888, + 0.0027204782236367464, + 0.0034599434584379196, + 0.017108574509620667, + 0.004271186888217926, + 0.022986706346273422, + 0.003049002029001713, + 0.004766796715557575, + 0.037695299834012985, + 0.00301948469132185, + 0.004983692429959774, + 0.006824590731412172, + 0.0038131489418447018, + 0.040553200989961624, + 0.005898905918002129, + 0.0027604340575635433, + 0.002022339031100273, + 0.002072324976325035, + 0.00313563016243279, + 0.003147661918774247, + 0.019432682543992996, + 0.0172028336673975, + 0.0032290539238601923, + 0.0014148305635899305, + 0.001972248312085867, + 0.0009859305573627353, + 0.0018457593396306038, + 0.0025813572574406862, + 0.0024442803114652634, + 0.001193121774122119, + 0.0012370777549222112, + 0.0015937060816213489, + 0.0019983684178441763, + 0.0018659038469195366, + 0.002167623257264495, + 0.0023796437308192253, + 0.0016891222912818193, + 0.002228411613032222, + 0.002220112131908536, + 0.006631312891840935, + 0.017147475853562355, + 0.009377172216773033, + 0.029867399483919144, + 0.010959390550851822, + 0.007819457910954952, + 0.02258467860519886, + 0.005470304749906063, + 0.01949460245668888, + 0.012661967426538467, + 0.010816835798323154, + 0.01163951214402914, + 0.008610828779637814, + 0.011820187792181969, + 0.008609239012002945, + 0.006826026830822229, + 0.014458732679486275, + 0.02682245336472988, + 0.011362113058567047, + 0.02388744428753853, + 0.02362617664039135, + 0.011723155155777931, + 0.02089592069387436, + 0.012552435509860516, + 0.023161111399531364, + 0.045777104794979095, + 0.013942080549895763, + 0.05252862349152565, + 0.040767740458250046, + 0.019083842635154724, + 0.027764396741986275, + 0.02249160222709179, + 0.019023628905415535, + 0.01921292394399643, + 0.017641868442296982, + 0.01239689625799656, + 0.021939218044281006, + 0.019980808719992638, + 0.020506761968135834, + 0.02674729749560356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.10035248100757599 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a static, tightly packed portrait-style close-up centered on a chest area?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a translucent bubble of magical energy with a delicate crystalline grid surface visible forming from fiery purple streams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a burst of searing crimson energy collide with the bubble, causing it to ripple, buckle, and twist before dissipating into fading points of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the revealed character a female warrior with raven-black hair, sharp grey eyes, and ornate dark steel armor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera smoothly tilt upward from the chest to reveal the warrior's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067780554294586 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067780554294586, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067780554294586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10222312808036804, + 0.23307588696479797, + 0.14553202688694, + 0.019095424562692642, + 0.034696392714977264, + 0.11020409315824509, + 0.17243723571300507, + 0.045378509908914566, + 0.06372000277042389, + 0.07363725453615189 + ], + "credit_map": [ + 0.006293695420026779, + 0.009982450865209103, + 0.00572249386459589, + 0.006991067435592413, + 0.025829195976257324, + 0.008054596371948719, + 0.03437605872750282, + 0.004973577335476875, + 0.009451717138290405, + 0.08428651839494705, + 0.006496287416666746, + 0.011196213774383068, + 0.014006352983415127, + 0.006343838758766651, + 0.08939650654792786, + 0.011898444965481758, + 0.004830281715840101, + 0.003234413219615817, + 0.0023905178532004356, + 0.013861503452062607, + 0.00799450185149908, + 0.04832756891846657, + 0.05772913619875908, + 0.0071641141548752785, + 0.0016308064805343747, + 0.0016674173530191183, + 0.0017278023296967149, + 0.003347405232489109, + 0.002716854214668274, + 0.0022930132690817118, + 0.0020359568297863007, + 0.003676171414554119, + 0.004048982635140419, + 0.0048148138448596, + 0.0038629737682640553, + 0.004168633837252855, + 0.00493343360722065, + 0.0030324961990118027, + 0.004135977476835251, + 0.005699084140360355, + 0.007147643249481916, + 0.01933181844651699, + 0.012596375308930874, + 0.030522437766194344, + 0.011692813597619534, + 0.008160188794136047, + 0.01606229692697525, + 0.004690523259341717, + 0.040915824472904205, + 0.020205169916152954, + 0.023286083713173866, + 0.022510740906000137, + 0.011114696972072124, + 0.02063530683517456, + 0.01871931552886963, + 0.015050108544528484, + 0.004350888542830944, + 0.009475460276007652, + 0.004000348970293999, + 0.006349868606775999, + 0.00735403411090374, + 0.0033286097459495068, + 0.006022215820848942, + 0.004497085697948933, + 0.011694022454321384, + 0.009584459476172924, + 0.0031794500537216663, + 0.01734820567071438, + 0.007293592672795057, + 0.003816188545897603, + 0.0062235016375780106, + 0.004580585286021233, + 0.005885134916752577, + 0.00785394012928009, + 0.007914533838629723, + 0.01196189783513546, + 0.010835916735231876, + 0.01000511646270752, + 0.010622401721775532, + 0.008558318018913269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.016067780554294586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_12/index_tensor([3591, 3591])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_12/index_tensor([3591, 3591])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..45fd0c5e3d9876d9db02199772a3c14a3e069e24 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_12/index_tensor([3591, 3591])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro close-up captured in stunning super-slow motion, executing a slow, sweeping trajectory, follows an oleander hawkmoth as it hovers with mesmerizing stability. The image reveals every minute detail of the creature's powerful wings, slowed to a hypnotic, balletic grace. The moth's torpedo-shaped body, covered in tiny, fuzzy olive-green hairs, remains almost perfectly still, while its intricate wings, frozen in time, display an intricate mosaic of olive green and khaki, cut through with sharp, geometric stripes of vibrant pink and soft white. The velvety texture of the wing scales is rendered with hyper-realistic clarity. As the camera follows the creature's subtle movements, the wings initiate a powerful downward motion, the leading edge cutting through the air and the tips twisting, revealing the complex aerodynamics at work. The background is a tapestry of dark, out-of-focus nighttime garden foliage, creating a beautiful bokeh that makes the moth's vivid colors pop. The high-speed nature photography, characterized by soft and directional lighting, captures the shimmering surface of the moth's large black eyes, casting subtle shadows that accentuate the three-dimensional forms of its body and wings, creating an atmosphere of pure curiosity focused on the mechanisms of nature.", + "seed": 18990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_12/index_3591_seed_18990_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03815971966832876, + "vlm_logprob": -0.03815971966832876, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.03815971966832876, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043431006371974945 + ], + "opposite_logprob_tokens": [ + -3.418431043624878 + ], + "logprob_reward": -0.043431006371974945, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.043431006371974945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26503533124923706, + 0.058629076927900314, + 0.13942067325115204, + 0.08851062506437302, + 0.07528933137655258, + 0.09637334942817688, + 0.0672832503914833, + 0.08446022868156433, + 0.0969320684671402, + 0.02806604467332363 + ], + "credit_map": [ + 0.021103616803884506, + 0.0056508868001401424, + 0.10951432585716248, + 0.012956028804183006, + 0.006581527646631002, + 0.011448106728494167, + 0.009802482090890408, + 0.08797837048768997, + 0.007476110477000475, + 0.00908272061496973, + 0.007988491095602512, + 0.005032671149820089, + 0.006413630209863186, + 0.0059565710835158825, + 0.010010396130383015, + 0.0066684880293905735, + 0.013326010666787624, + 0.01838902197778225, + 0.016115743666887283, + 0.011659238487482071, + 0.037634897977113724, + 0.016597948968410492, + 0.015801258385181427, + 0.009896557778120041, + 0.012098806910216808, + 0.011341113597154617, + 0.011101249605417252, + 0.010417145676910877, + 0.0081490408629179, + 0.0072407484985888, + 0.011880354955792427, + 0.01628216728568077, + 0.007365784142166376, + 0.010907599702477455, + 0.010133622214198112, + 0.009025059640407562, + 0.006091824732720852, + 0.008071775548160076, + 0.012695151381194592, + 0.010998514480888844, + 0.007222667336463928, + 0.013377390801906586, + 0.012464995495975018, + 0.010056358762085438, + 0.01570267789065838, + 0.016314420849084854, + 0.011495036073029041, + 0.009739799425005913, + 0.007763873320072889, + 0.010918860323727131, + 0.010570726357400417, + 0.006775770336389542, + 0.0096063744276762, + 0.004842274356633425, + 0.008150369860231876, + 0.00865500420331955, + 0.011937654577195644, + 0.010497753508388996, + 0.011607857421040535, + 0.007813293486833572, + 0.0078084273263812065, + 0.010033372789621353, + 0.013900727033615112, + 0.01086114440113306, + 0.009667654521763325, + 0.012642743065953255, + 0.013202181085944176, + 0.008193091489374638, + 0.012771380133926868, + 0.017656810581684113, + 0.011179273948073387, + 0.011618923395872116, + 0.0035254559479653835, + 0.003994716331362724, + 0.0038203380536288023, + 0.0029222110752016306, + 0.0039055170491337776, + 0.003436689032241702, + 0.0035615714732557535, + 0.002899543847888708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.043431006371974945 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007655922323465347 + ], + "opposite_logprob_tokens": [ + -4.882656097412109 + ], + "logprob_reward": -0.007655922323465347, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007655922323465347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19500695168972015, + 0.08164099603891373, + 0.13897986710071564, + 0.05286433920264244, + 0.21043384075164795, + 0.06344512850046158, + 0.101780965924263, + 0.03882169723510742, + 0.08184070885181427, + 0.0351855531334877 + ], + "credit_map": [ + 0.01808755286037922, + 0.016970627009868622, + 0.04851444065570831, + 0.02104875259101391, + 0.016575470566749573, + 0.015782266855239868, + 0.015334950760006905, + 0.042692895978689194, + 0.010556032881140709, + 0.011093959212303162, + 0.013424471952021122, + 0.008684399537742138, + 0.009177925996482372, + 0.006795342545956373, + 0.0094490060582757, + 0.01245985645800829, + 0.02159309573471546, + 0.014681034721434116, + 0.017489276826381683, + 0.013711061328649521, + 0.025125334039330482, + 0.01275752391666174, + 0.02050582878291607, + 0.013116712681949139, + 0.007002244703471661, + 0.006781864911317825, + 0.006287718657404184, + 0.005878872238099575, + 0.005193834193050861, + 0.0057569160126149654, + 0.009480603970587254, + 0.006482283119112253, + 0.0181938037276268, + 0.029943639412522316, + 0.03367646783590317, + 0.020865386351943016, + 0.03104688599705696, + 0.01580013893544674, + 0.02685842104256153, + 0.034049101173877716, + 0.004380342084914446, + 0.011524764820933342, + 0.008677263744175434, + 0.0057990700006484985, + 0.006062785163521767, + 0.0077772377990186214, + 0.010166211985051632, + 0.009057450108230114, + 0.015326841734349728, + 0.013095682486891747, + 0.016638198867440224, + 0.010464848019182682, + 0.00856901053339243, + 0.010797165334224701, + 0.01361914537847042, + 0.013270072638988495, + 0.006051256787031889, + 0.004706540144979954, + 0.005422740709036589, + 0.004806842654943466, + 0.002656168770045042, + 0.0036765849217772484, + 0.007750642020255327, + 0.0037509233225136995, + 0.007578244898468256, + 0.012176977470517159, + 0.010709107853472233, + 0.0068032084964215755, + 0.02004971168935299, + 0.00804628524929285, + 0.009161263704299927, + 0.0073159076273441315, + 0.0034573348239064217, + 0.004433023743331432, + 0.004514798521995544, + 0.005014901515096426, + 0.0037646975833922625, + 0.004906465765088797, + 0.0033312137238681316, + 0.005763113498687744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.007655922323465347 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308246821165085 + ], + "opposite_logprob_tokens": [ + -3.1680824756622314 + ], + "logprob_reward": -0.04308246821165085, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308246821165085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12145344167947769, + 0.0696307122707367, + 0.08127683401107788, + 0.11198326200246811, + 0.21954327821731567, + 0.06592295318841934, + 0.17572180926799774, + 0.044542644172906876, + 0.06256121397018433, + 0.04736390337347984 + ], + "credit_map": [ + 0.012276522815227509, + 0.011525479145348072, + 0.034857288002967834, + 0.010576212778687477, + 0.00847479235380888, + 0.008026117458939552, + 0.00820641778409481, + 0.027510611340403557, + 0.01097141019999981, + 0.007635338231921196, + 0.013165193609893322, + 0.006208963692188263, + 0.005370429717004299, + 0.006196083966642618, + 0.010614467784762383, + 0.009468828327953815, + 0.010943567380309105, + 0.007940657436847687, + 0.01102453749626875, + 0.0072273509576916695, + 0.011698289774358273, + 0.00919128768146038, + 0.011709493584930897, + 0.011541648767888546, + 0.0211174376308918, + 0.013617771677672863, + 0.014491381123661995, + 0.011059964075684547, + 0.011084162630140781, + 0.009886822663247585, + 0.014544767327606678, + 0.01618095114827156, + 0.01448045950382948, + 0.03618962690234184, + 0.029783986508846283, + 0.024560796096920967, + 0.02843172289431095, + 0.018843958154320717, + 0.03371139615774155, + 0.033541325479745865, + 0.006871076766401529, + 0.005837313365191221, + 0.00943808350712061, + 0.005132655147463083, + 0.005263722967356443, + 0.005458673927932978, + 0.01285452488809824, + 0.015066901221871376, + 0.01639113947749138, + 0.022264759987592697, + 0.023562129586935043, + 0.029885979369282722, + 0.01554595772176981, + 0.015619381330907345, + 0.02836138755083084, + 0.02409106306731701, + 0.006884596310555935, + 0.005072626750916243, + 0.0068384697660803795, + 0.004595807753503323, + 0.0033616533037275076, + 0.0053014205768704414, + 0.006804741453379393, + 0.005683323368430138, + 0.007160552777349949, + 0.008164807222783566, + 0.008065748028457165, + 0.005384155083447695, + 0.013717953115701675, + 0.005938361398875713, + 0.007995941676199436, + 0.006133692804723978, + 0.005087363533675671, + 0.006950369570404291, + 0.006474131718277931, + 0.0059897019527852535, + 0.004726243205368519, + 0.007418840657919645, + 0.005982355214655399, + 0.004734892398118973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.04308246821165085 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625400088727474 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625400088727474, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625400088727474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09226257354021072, + 0.07787452638149261, + 0.09099553525447845, + 0.1931418627500534, + 0.056050416082143784, + 0.03177342936396599, + 0.34565746784210205, + 0.07425922900438309, + 0.021039694547653198, + 0.01694522611796856 + ], + "credit_map": [ + 0.011021211743354797, + 0.010662849992513657, + 0.014927469193935394, + 0.014195759780704975, + 0.009892357513308525, + 0.008565650321543217, + 0.010060661472380161, + 0.012936608865857124, + 0.010469893924891949, + 0.009322572499513626, + 0.013722117058932781, + 0.007996432483196259, + 0.008511869236826897, + 0.008239985443651676, + 0.01252075657248497, + 0.0070908949710428715, + 0.024832185357809067, + 0.008861290290951729, + 0.016630355268716812, + 0.007322799414396286, + 0.009032990783452988, + 0.005767828784883022, + 0.00674611097201705, + 0.011801979504525661, + 0.024958433583378792, + 0.03173799440264702, + 0.02548898383975029, + 0.027630919590592384, + 0.01851707510650158, + 0.017895597964525223, + 0.02651785872876644, + 0.02039499208331108, + 0.005396494176238775, + 0.010630292817950249, + 0.008682683110237122, + 0.005648946389555931, + 0.005594675429165363, + 0.004237810615450144, + 0.0075004491955041885, + 0.008359061554074287, + 0.0026072801556438208, + 0.005175720900297165, + 0.004183616489171982, + 0.002863355679437518, + 0.00404697610065341, + 0.0039190128445625305, + 0.00466949725523591, + 0.004307969007641077, + 0.05855879560112953, + 0.04266901686787605, + 0.05814963951706886, + 0.027412625029683113, + 0.03006003610789776, + 0.041349396109580994, + 0.04220946505665779, + 0.04524851217865944, + 0.012969431467354298, + 0.01020027045160532, + 0.011405827477574348, + 0.008102135732769966, + 0.006119695026427507, + 0.0059998491778969765, + 0.012193846516311169, + 0.0072681717574596405, + 0.003645027056336403, + 0.002367110690101981, + 0.0032428386621177197, + 0.0019223348936066031, + 0.0022899627219885588, + 0.0019843238405883312, + 0.00222789472900331, + 0.0033602023031562567, + 0.002187614794820547, + 0.0023524395655840635, + 0.0021870294585824013, + 0.0016386342467740178, + 0.0015139764873310924, + 0.002927132649347186, + 0.002240665489807725, + 0.0018977333093062043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.007625400088727474 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890038013458252 + ], + "opposite_logprob_tokens": [ + -2.464003801345825 + ], + "logprob_reward": -0.0890038013458252, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890038013458252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3521795868873596, + 0.051581401377916336, + 0.0734226331114769, + 0.16930674016475677, + 0.06599381566047668, + 0.04021858051419258, + 0.14066864550113678, + 0.035717178136110306, + 0.02285437285900116, + 0.04805702716112137 + ], + "credit_map": [ + 0.03182005509734154, + 0.019221749156713486, + 0.1119980663061142, + 0.027301371097564697, + 0.02087644673883915, + 0.020638007670640945, + 0.03081430122256279, + 0.08950959146022797, + 0.007159887813031673, + 0.005566183943301439, + 0.010514120571315289, + 0.004793065600097179, + 0.005596541799604893, + 0.00431414507329464, + 0.006361440755426884, + 0.007276017684489489, + 0.007923067547380924, + 0.010292050428688526, + 0.008707050234079361, + 0.003886683378368616, + 0.018625257536768913, + 0.008830518461763859, + 0.007484909147024155, + 0.007673096377402544, + 0.03142070025205612, + 0.02266557700932026, + 0.023078767582774162, + 0.01641690544784069, + 0.018941372632980347, + 0.014823349192738533, + 0.0239089485257864, + 0.018051117658615112, + 0.005993400700390339, + 0.011524816043674946, + 0.008241766132414341, + 0.004977669101208448, + 0.013504835776984692, + 0.006796012166887522, + 0.008048253133893013, + 0.006907064467668533, + 0.0031307374592870474, + 0.004408117383718491, + 0.005673832260072231, + 0.0040075951255857944, + 0.00514738354831934, + 0.00450481241568923, + 0.0063100531697273254, + 0.007036049384623766, + 0.014852997846901417, + 0.01598263345658779, + 0.02327047660946846, + 0.012166056782007217, + 0.017048168927431107, + 0.017949512228369713, + 0.021036159247159958, + 0.018362633883953094, + 0.004525552503764629, + 0.005985461641103029, + 0.005226890090852976, + 0.0035883616656064987, + 0.0025795770343393087, + 0.0032654679380357265, + 0.006105718202888966, + 0.004440148826688528, + 0.003752922173589468, + 0.004052634350955486, + 0.002946895780041814, + 0.002137674717232585, + 0.0023019928485155106, + 0.0026164802256971598, + 0.0027308196295052767, + 0.0023149538319557905, + 0.004606346134096384, + 0.005431933794170618, + 0.006206005811691284, + 0.007819180376827717, + 0.004684041254222393, + 0.007710373029112816, + 0.0068815634585917, + 0.004717587027698755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.0890038013458252 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up captured in stunning super-slow motion, executing a slow, sweeping trajectory, follows an oleander hawkmoth as it hovers with mesmerizing stability. The image reveals every minute detail of the creature's powerful wings, slowed to a hypnotic, balletic grace. The moth's torpedo-shaped body, covered in tiny, fuzzy olive-green hairs, remains almost perfectly still, while its intricate wings, frozen in time, display an intricate mosaic of olive green and khaki, cut through with sharp, geometric stripes of vibrant pink and soft white. The velvety texture of the wing scales is rendered with hyper-realistic clarity. As the camera follows the creature's subtle movements, the wings initiate a powerful downward motion, the leading edge cutting through the air and the tips twisting, revealing the complex aerodynamics at work. The background is a tapestry of dark, out-of-focus nighttime garden foliage, creating a beautiful bokeh that makes the moth's vivid colors pop. The high-speed nature photography, characterized by soft and directional lighting, captures the shimmering surface of the moth's large black eyes, casting subtle shadows that accentuate the three-dimensional forms of its body and wings, creating an atmosphere of pure curiosity focused on the mechanisms of nature.", + "seed": 18991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_12/index_3591_seed_18991_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04188642036169767, + "vlm_logprob": -0.04188642036169767, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.04188642036169767, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08979971706867218 + ], + "opposite_logprob_tokens": [ + -2.589799642562866 + ], + "logprob_reward": -0.08979971706867218, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08979971706867218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15630276501178741, + 0.12359301000833511, + 0.14858175814151764, + 0.02701987698674202, + 0.14976774156093597, + 0.09386197477579117, + 0.09570740908384323, + 0.04137882962822914, + 0.10072796046733856, + 0.06305868178606033 + ], + "credit_map": [ + 0.01632564142346382, + 0.017695149406790733, + 0.01577720418572426, + 0.01346251554787159, + 0.0270171370357275, + 0.024400973692536354, + 0.016684524714946747, + 0.02493961714208126, + 0.0147702069953084, + 0.022180737927556038, + 0.00933013204485178, + 0.010248040780425072, + 0.025292880833148956, + 0.017536530271172523, + 0.008737361989915371, + 0.01549711637198925, + 0.028937363997101784, + 0.011748685501515865, + 0.015362490899860859, + 0.019845059141516685, + 0.021145083010196686, + 0.016634156927466393, + 0.018627043813467026, + 0.016281859949231148, + 0.0029281722381711006, + 0.0020424453541636467, + 0.0019129522843286395, + 0.0015215898165479302, + 0.0043628583662211895, + 0.004431269131600857, + 0.002634502248838544, + 0.007186088245362043, + 0.03534587100148201, + 0.011121580377221107, + 0.0417841300368309, + 0.013647619634866714, + 0.0128168985247612, + 0.010075363330543041, + 0.00785213802009821, + 0.01712413690984249, + 0.009141608141362667, + 0.010247120633721352, + 0.009445982053875923, + 0.01368054561316967, + 0.01318990159779787, + 0.013753654435276985, + 0.011592292226850986, + 0.012810865417122841, + 0.0061561898328363895, + 0.0138179836794734, + 0.011037841439247131, + 0.013094485737383366, + 0.013678538613021374, + 0.014559135772287846, + 0.009676472283899784, + 0.013686763122677803, + 0.004902761895209551, + 0.0036891165655106306, + 0.003170931711792946, + 0.004953501280397177, + 0.0076997787691652775, + 0.00636937515810132, + 0.004419771954417229, + 0.006173588801175356, + 0.013238196261227131, + 0.013068966567516327, + 0.005885452497750521, + 0.013637898489832878, + 0.016480950638651848, + 0.014683871529996395, + 0.010795003734529018, + 0.01293762493878603, + 0.010552126914262772, + 0.006379441358149052, + 0.008035515435039997, + 0.007693935185670853, + 0.006120427045971155, + 0.008729151450097561, + 0.007265222724527121, + 0.008282868191599846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.08979971706867218 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008671131916344166 + ], + "opposite_logprob_tokens": [ + -4.758671283721924 + ], + "logprob_reward": -0.008671131916344166, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008671131916344166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09409799426794052, + 0.0502198301255703, + 0.12979276478290558, + 0.05855207517743111, + 0.2295577973127365, + 0.03974694386124611, + 0.10763459652662277, + 0.05459972098469734, + 0.16566869616508484, + 0.0701296404004097 + ], + "credit_map": [ + 0.009532764554023743, + 0.011208507232367992, + 0.009595625102519989, + 0.00894698966294527, + 0.014578907750546932, + 0.014323428273200989, + 0.008424106054008007, + 0.0174876619130373, + 0.007321878336369991, + 0.008565817028284073, + 0.004693140275776386, + 0.005972383078187704, + 0.008307811804115772, + 0.006660403683781624, + 0.0040791695937514305, + 0.004619222600013018, + 0.03871569409966469, + 0.01230813842266798, + 0.01701204665005207, + 0.011407990008592606, + 0.016181733459234238, + 0.012392844073474407, + 0.01124246884137392, + 0.010531836189329624, + 0.006137087941169739, + 0.003932953346520662, + 0.005536997225135565, + 0.004204310476779938, + 0.007743224035948515, + 0.008338117972016335, + 0.006026010494679213, + 0.01663336530327797, + 0.06590981781482697, + 0.013531377539038658, + 0.08381091803312302, + 0.011154347099363804, + 0.01273777149617672, + 0.007961167022585869, + 0.005565687548369169, + 0.028886718675494194, + 0.0039381589740514755, + 0.004304794128984213, + 0.003842436708509922, + 0.005771568510681391, + 0.005868989508599043, + 0.00621568039059639, + 0.004323817323893309, + 0.005481497384607792, + 0.00909118540585041, + 0.01151107344776392, + 0.011613414622843266, + 0.014609244652092457, + 0.010302554816007614, + 0.023624945431947708, + 0.010675385594367981, + 0.016206782311201096, + 0.007099356036633253, + 0.006034988444298506, + 0.0047008562833070755, + 0.00489128427579999, + 0.011908681131899357, + 0.008206277154386044, + 0.004445235710591078, + 0.007313041482120752, + 0.017669742926955223, + 0.014332802966237068, + 0.012403537519276142, + 0.02222943864762783, + 0.026562657207250595, + 0.02461358532309532, + 0.02136390283703804, + 0.026493007317185402, + 0.008157521486282349, + 0.00618972210213542, + 0.005359635688364506, + 0.008605212904512882, + 0.007953304797410965, + 0.012847554869949818, + 0.011305268853902817, + 0.00971140991896391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.008671131916344166 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982204221189022 + ], + "opposite_logprob_tokens": [ + -3.5298221111297607 + ], + "logprob_reward": -0.02982204221189022, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982204221189022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19150373339653015, + 0.08494924753904343, + 0.10730783641338348, + 0.03464260324835777, + 0.11512725800275803, + 0.028407586738467216, + 0.0673188716173172, + 0.055796947330236435, + 0.2328411191701889, + 0.08210477977991104 + ], + "credit_map": [ + 0.02348124422132969, + 0.02055364102125168, + 0.026514841243624687, + 0.017617862671613693, + 0.02598937228322029, + 0.029411017894744873, + 0.015222408808767796, + 0.03271334990859032, + 0.007933984510600567, + 0.011331678368151188, + 0.007854266092181206, + 0.009020751342177391, + 0.01669221930205822, + 0.01520646084100008, + 0.007033877540379763, + 0.009876019321382046, + 0.022505493834614754, + 0.006066945381462574, + 0.011613987386226654, + 0.012153904885053635, + 0.012768407352268696, + 0.01328346598893404, + 0.013875998556613922, + 0.01503963302820921, + 0.003440035739913583, + 0.004785523284226656, + 0.002876889193430543, + 0.003293206449598074, + 0.004841484595090151, + 0.00455082580447197, + 0.005426098592579365, + 0.005428540986031294, + 0.02536921761929989, + 0.007907964289188385, + 0.0261517446488142, + 0.009965233504772186, + 0.009534587152302265, + 0.01280726958066225, + 0.008070834912359715, + 0.015320404432713985, + 0.001874852110631764, + 0.006375330034643412, + 0.003648026380687952, + 0.004320876207202673, + 0.004209480248391628, + 0.002588427858427167, + 0.0018883843440562487, + 0.0035022099036723375, + 0.005105482414364815, + 0.007538252975791693, + 0.010446149855852127, + 0.011494908481836319, + 0.007732894271612167, + 0.009220677427947521, + 0.0060859788209199905, + 0.009694525972008705, + 0.00602124584838748, + 0.00703330896794796, + 0.004019258078187704, + 0.007191698998212814, + 0.009506062604486942, + 0.008494003675878048, + 0.005453053861856461, + 0.008078312501311302, + 0.0282752588391304, + 0.028330709785223007, + 0.016428295522928238, + 0.03511849790811539, + 0.035729460418224335, + 0.03288894519209862, + 0.02507176622748375, + 0.03099817782640457, + 0.010998099111020565, + 0.007137914653867483, + 0.005804779008030891, + 0.013294470496475697, + 0.010257708840072155, + 0.013525175862014294, + 0.008702484890818596, + 0.012384137138724327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.02982204221189022 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809259459376335 + ], + "opposite_logprob_tokens": [ + -3.288092613220215 + ], + "logprob_reward": -0.03809259459376335, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809259459376335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12927588820457458, + 0.05000066012144089, + 0.034575726836919785, + 0.07955198734998703, + 0.057021331042051315, + 0.015506655909121037, + 0.23994751274585724, + 0.0907912477850914, + 0.032638221979141235, + 0.2706908583641052 + ], + "credit_map": [ + 0.016865568235516548, + 0.013396061956882477, + 0.025304557755589485, + 0.011345651932060719, + 0.014512622728943825, + 0.019108561798930168, + 0.01143112313002348, + 0.01731172204017639, + 0.007988695055246353, + 0.008994979783892632, + 0.005832701455801725, + 0.0051314132288098335, + 0.00689173536375165, + 0.006802849471569061, + 0.003647630335763097, + 0.0047106496058404446, + 0.007443658076226711, + 0.0028718498069792986, + 0.003924361430108547, + 0.0034054501447826624, + 0.003997101914137602, + 0.003727453527972102, + 0.004332108423113823, + 0.004873739555478096, + 0.00872703455388546, + 0.00594807043671608, + 0.007815352641046047, + 0.00450551463291049, + 0.012795730493962765, + 0.010208326391875744, + 0.007106697652488947, + 0.022445248439908028, + 0.0104912668466568, + 0.0049669137224555016, + 0.00959215871989727, + 0.004860305227339268, + 0.006452904082834721, + 0.007689025718718767, + 0.0053289104253053665, + 0.007639838848263025, + 0.0018688322743400931, + 0.002031531184911728, + 0.0016250238986685872, + 0.0024439534172415733, + 0.0017670777160674334, + 0.002204853342846036, + 0.0014211685629561543, + 0.002144214231520891, + 0.020295320078730583, + 0.020015859976410866, + 0.050729166716337204, + 0.02722746692597866, + 0.023863399401307106, + 0.04214413836598396, + 0.020412251353263855, + 0.03525988385081291, + 0.016952188685536385, + 0.010122929699718952, + 0.007808622904121876, + 0.008042462170124054, + 0.012687638401985168, + 0.013762176968157291, + 0.006781943142414093, + 0.014633282087743282, + 0.003203843953087926, + 0.003824049374088645, + 0.0030808253213763237, + 0.003047569189220667, + 0.0033316488843411207, + 0.005995393730700016, + 0.004010946024209261, + 0.006143942940980196, + 0.029127664864063263, + 0.030062954872846603, + 0.024559153243899345, + 0.032564375549554825, + 0.03032951056957245, + 0.04876619949936867, + 0.039589133113622665, + 0.03569183498620987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.03809259459376335 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304661601781845 + ], + "opposite_logprob_tokens": [ + -3.168046712875366 + ], + "logprob_reward": -0.04304661601781845, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304661601781845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25160470604896545, + 0.2548297643661499, + 0.0537235327064991, + 0.055829890072345734, + 0.12129884213209152, + 0.029052192345261574, + 0.05919039621949196, + 0.021791521459817886, + 0.10399996489286423, + 0.048679154366254807 + ], + "credit_map": [ + 0.030047640204429626, + 0.0253650713711977, + 0.03617382049560547, + 0.0211898535490036, + 0.031284574419260025, + 0.037070129066705704, + 0.02269975282251835, + 0.047773852944374084, + 0.03404948487877846, + 0.047004617750644684, + 0.022287368774414062, + 0.02209121733903885, + 0.04399759694933891, + 0.035213299095630646, + 0.01997894048690796, + 0.030207229778170586, + 0.010285241529345512, + 0.005930066108703613, + 0.004541588947176933, + 0.005553639493882656, + 0.0068993158638477325, + 0.006868066731840372, + 0.006366168148815632, + 0.007279445882886648, + 0.006837110500782728, + 0.005219016224145889, + 0.003578413277864456, + 0.0032654593233019114, + 0.008827021345496178, + 0.009167821146547794, + 0.00452036689966917, + 0.014414679259061813, + 0.028992140665650368, + 0.0068303742446005344, + 0.03932579606771469, + 0.007023856043815613, + 0.01046860683709383, + 0.008121998980641365, + 0.005283590406179428, + 0.015252477489411831, + 0.0027152562979608774, + 0.003624681616201997, + 0.0036086023319512606, + 0.005391268990933895, + 0.0029603124130517244, + 0.0033306945115327835, + 0.003135259263217449, + 0.004286116920411587, + 0.0046778228133916855, + 0.005414940882474184, + 0.009115743450820446, + 0.006128894630819559, + 0.006550670135766268, + 0.013284314423799515, + 0.005087690893560648, + 0.008930320851504803, + 0.0032348069362342358, + 0.002516654320061207, + 0.0016174712218344212, + 0.002016331534832716, + 0.0039208680391311646, + 0.003242061473429203, + 0.0018891510553658009, + 0.0033541761804372072, + 0.015004387125372887, + 0.011552081443369389, + 0.006956701632589102, + 0.016639672219753265, + 0.014858857728540897, + 0.016339564695954323, + 0.009562445804476738, + 0.013086245395243168, + 0.006270165555179119, + 0.004773040767759085, + 0.005533342715352774, + 0.005637213122099638, + 0.0044134315103292465, + 0.007377904839813709, + 0.00806121714413166, + 0.006612840574234724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.04304661601781845 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up captured in stunning super-slow motion, executing a slow, sweeping trajectory, follows an oleander hawkmoth as it hovers with mesmerizing stability. The image reveals every minute detail of the creature's powerful wings, slowed to a hypnotic, balletic grace. The moth's torpedo-shaped body, covered in tiny, fuzzy olive-green hairs, remains almost perfectly still, while its intricate wings, frozen in time, display an intricate mosaic of olive green and khaki, cut through with sharp, geometric stripes of vibrant pink and soft white. The velvety texture of the wing scales is rendered with hyper-realistic clarity. As the camera follows the creature's subtle movements, the wings initiate a powerful downward motion, the leading edge cutting through the air and the tips twisting, revealing the complex aerodynamics at work. The background is a tapestry of dark, out-of-focus nighttime garden foliage, creating a beautiful bokeh that makes the moth's vivid colors pop. The high-speed nature photography, characterized by soft and directional lighting, captures the shimmering surface of the moth's large black eyes, casting subtle shadows that accentuate the three-dimensional forms of its body and wings, creating an atmosphere of pure curiosity focused on the mechanisms of nature.", + "seed": 18992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_12/index_3591_seed_18992_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05629437956959009, + "vlm_logprob": -0.05629437956959009, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.05629437956959009, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05718559771776199 + ], + "opposite_logprob_tokens": [ + -3.057185649871826 + ], + "logprob_reward": -0.05718559771776199, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05718559771776199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09068865329027176, + 0.13051387667655945, + 0.2059476673603058, + 0.07053129374980927, + 0.15095539391040802, + 0.05365477874875069, + 0.13669955730438232, + 0.08947927504777908, + 0.03020547516644001, + 0.04132402688264847 + ], + "credit_map": [ + 0.01820998266339302, + 0.014827742241322994, + 0.011311224661767483, + 0.010459290817379951, + 0.00606730580329895, + 0.006705213338136673, + 0.007513847202062607, + 0.015594043768942356, + 0.009975383058190346, + 0.004905706271529198, + 0.04807879403233528, + 0.0063053518533706665, + 0.004069164860993624, + 0.005069696810096502, + 0.0034063055645674467, + 0.04870348423719406, + 0.014628770761191845, + 0.009974689222872257, + 0.07491673529148102, + 0.010058904066681862, + 0.006800870411098003, + 0.010117597877979279, + 0.007556795142591, + 0.07189331948757172, + 0.010335225611925125, + 0.007991973310709, + 0.009977986104786396, + 0.011217585764825344, + 0.006482570432126522, + 0.006272746250033379, + 0.009696006774902344, + 0.008557198569178581, + 0.0161896962672472, + 0.01929231360554695, + 0.02587733417749405, + 0.01711248606443405, + 0.014916401356458664, + 0.016974087804555893, + 0.02103918045759201, + 0.0195538979023695, + 0.007743079215288162, + 0.0049811373464763165, + 0.007888203486800194, + 0.007576913107186556, + 0.004659738857299089, + 0.005135627929121256, + 0.007675754372030497, + 0.007994324900209904, + 0.007795260287821293, + 0.004418819211423397, + 0.05048881098628044, + 0.007208984810858965, + 0.005605889949947596, + 0.006235959008336067, + 0.005512951873242855, + 0.04943288117647171, + 0.00718088960275054, + 0.006908482871949673, + 0.018077369779348373, + 0.017021799460053444, + 0.012866172939538956, + 0.006926216650754213, + 0.007118196226656437, + 0.013380148448050022, + 0.004665323067456484, + 0.0033014139626175165, + 0.0036811260506510735, + 0.0031683791894465685, + 0.0033213282003998756, + 0.003938435576856136, + 0.0031435489654541016, + 0.0049859220162034035, + 0.004937686957418919, + 0.007932748645544052, + 0.004835488274693489, + 0.0042058187536895275, + 0.005474150646477938, + 0.004092047922313213, + 0.0049153589643538, + 0.004930730443447828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.05718559771776199 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020607180893421173 + ], + "opposite_logprob_tokens": [ + -3.8956072330474854 + ], + "logprob_reward": -0.020607180893421173, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020607180893421173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10749179124832153, + 0.32312342524528503, + 0.07321979850530624, + 0.07538307458162308, + 0.09200829267501831, + 0.04177599400281906, + 0.08218450844287872, + 0.1291559636592865, + 0.026637939736247063, + 0.049019213765859604 + ], + "credit_map": [ + 0.011530320160090923, + 0.024426208809018135, + 0.012412049807608128, + 0.013688149861991405, + 0.011149829253554344, + 0.010542604140937328, + 0.011413638480007648, + 0.012328979559242725, + 0.033217903226614, + 0.01144429948180914, + 0.11823002994060516, + 0.01286489050835371, + 0.011870251037180424, + 0.011572383344173431, + 0.009341802448034286, + 0.11458183079957962, + 0.0054311431013047695, + 0.004511727951467037, + 0.021212458610534668, + 0.004051722586154938, + 0.0038691593799740076, + 0.005224392283707857, + 0.006648771930485964, + 0.02227042429149151, + 0.018158061429858208, + 0.009515817277133465, + 0.009398485533893108, + 0.00861024297773838, + 0.0052495915442705154, + 0.006271399091929197, + 0.007545873057097197, + 0.010633600875735283, + 0.016878467053174973, + 0.02009063959121704, + 0.008654189296066761, + 0.007768835406750441, + 0.010819291695952415, + 0.008685136213898659, + 0.008357553742825985, + 0.010754182003438473, + 0.006635237485170364, + 0.004971649032086134, + 0.0069791944697499275, + 0.006541289854794741, + 0.003386625088751316, + 0.003648864571005106, + 0.0036236699670553207, + 0.005989460274577141, + 0.0069868979044258595, + 0.006231543608009815, + 0.020434044301509857, + 0.007341319695115089, + 0.005139461252838373, + 0.006684533786028624, + 0.0071421535685658455, + 0.0222245454788208, + 0.008303332142531872, + 0.014025515876710415, + 0.021499643102288246, + 0.012583763338625431, + 0.016848597675561905, + 0.021938243880867958, + 0.009035593830049038, + 0.024921264499425888, + 0.002416793256998062, + 0.0050481390208005905, + 0.0032456167973577976, + 0.0030334254261106253, + 0.0032467220444232225, + 0.003465107409283519, + 0.0025582651142030954, + 0.003623869502916932, + 0.006354549899697304, + 0.00804806686937809, + 0.007069685496389866, + 0.005744878202676773, + 0.005059920717030764, + 0.004921299405395985, + 0.005943345371633768, + 0.00587746687233448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.020607180893421173 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811485692858696 + ], + "opposite_logprob_tokens": [ + -3.2881147861480713 + ], + "logprob_reward": -0.03811485692858696, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811485692858696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10157155990600586, + 0.2452445924282074, + 0.07011137157678604, + 0.09609314799308777, + 0.0898231491446495, + 0.06147044897079468, + 0.09158922731876373, + 0.14520861208438873, + 0.06229063495993614, + 0.036597270518541336 + ], + "credit_map": [ + 0.012805670499801636, + 0.019128765910863876, + 0.012243351899087429, + 0.014363406226038933, + 0.009600190445780754, + 0.008760425262153149, + 0.010944442823529243, + 0.013725302182137966, + 0.014040573500096798, + 0.009986490942537785, + 0.08912841230630875, + 0.012423677369952202, + 0.011799335479736328, + 0.009698054753243923, + 0.0091688958927989, + 0.08899913728237152, + 0.007231442257761955, + 0.006162463221698999, + 0.012859975919127464, + 0.005783921107649803, + 0.007445249240845442, + 0.005516837816685438, + 0.010909497737884521, + 0.014201977290213108, + 0.010891244746744633, + 0.015486997552216053, + 0.01042159367352724, + 0.013368654064834118, + 0.011457034386694431, + 0.0100924177095294, + 0.009306688793003559, + 0.015068508684635162, + 0.015920432284474373, + 0.01024363562464714, + 0.00790978129953146, + 0.01006624847650528, + 0.011493680998682976, + 0.010945158079266548, + 0.013240349479019642, + 0.010003848932683468, + 0.007116667926311493, + 0.007458704058080912, + 0.007693702820688486, + 0.008767473511397839, + 0.00646639708429575, + 0.006173902191221714, + 0.0093716811388731, + 0.008421916514635086, + 0.010640185326337814, + 0.007557850796729326, + 0.01493555773049593, + 0.013463045470416546, + 0.010440373793244362, + 0.008296572603285313, + 0.00951109454035759, + 0.016744546592235565, + 0.011281133629381657, + 0.012982035987079144, + 0.022635620087385178, + 0.0253642238676548, + 0.016092581674456596, + 0.021257827058434486, + 0.013660243712365627, + 0.021934935823082924, + 0.006247811950743198, + 0.004016020800918341, + 0.014082253910601139, + 0.004665559157729149, + 0.006660697050392628, + 0.008876937441527843, + 0.005108947865664959, + 0.012632397934794426, + 0.003785841166973114, + 0.0034966038074344397, + 0.006807335652410984, + 0.0036600467283278704, + 0.0031078632455319166, + 0.006989825051277876, + 0.004003089386969805, + 0.004746661987155676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.03811485692858696 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249881185591221 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249881185591221, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249881185591221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1520301103591919, + 0.33647263050079346, + 0.061764705926179886, + 0.03351161256432533, + 0.05964915081858635, + 0.16159449517726898, + 0.11943414807319641, + 0.017170920968055725, + 0.04527393728494644, + 0.013098357245326042 + ], + "credit_map": [ + 0.028673220425844193, + 0.021336330100893974, + 0.01649203896522522, + 0.023171253502368927, + 0.01010134257376194, + 0.012462817132472992, + 0.012443951331079006, + 0.02734915353357792, + 0.016064602881669998, + 0.008428648114204407, + 0.13805852830410004, + 0.010875373147428036, + 0.008222008123993874, + 0.009721336886286736, + 0.006942123640328646, + 0.13816003501415253, + 0.0046326476149261, + 0.003762700129300356, + 0.020658224821090698, + 0.004114443436264992, + 0.002191849984228611, + 0.0031637654174119234, + 0.0026089935563504696, + 0.02063208445906639, + 0.0041785710491240025, + 0.004931724164634943, + 0.0038408986292779446, + 0.004932726267725229, + 0.0041205729357898235, + 0.003477354533970356, + 0.0035432567819952965, + 0.004486508201807737, + 0.006210250314325094, + 0.009328262880444527, + 0.004250964615494013, + 0.006825104355812073, + 0.006974967196583748, + 0.00945869367569685, + 0.009093442000448704, + 0.007507462054491043, + 0.0215920340269804, + 0.01583862490952015, + 0.0317823626101017, + 0.032794803380966187, + 0.009157351218163967, + 0.011074707843363285, + 0.014167233370244503, + 0.025187375023961067, + 0.021806539967656136, + 0.005888409446924925, + 0.033591095358133316, + 0.007054529618471861, + 0.0058648292906582355, + 0.0050325337797403336, + 0.004955907352268696, + 0.03524031490087509, + 0.0017826864495873451, + 0.0018307974096387625, + 0.0025549635756760836, + 0.0024659933988004923, + 0.002336595207452774, + 0.002185310237109661, + 0.0012455199612304568, + 0.002769054612144828, + 0.004941032733768225, + 0.003344259923323989, + 0.00887388363480568, + 0.004487012978643179, + 0.005451660603284836, + 0.005744319409132004, + 0.004068060778081417, + 0.008363707922399044, + 0.0016295521054416895, + 0.0023239227011799812, + 0.0019074868177995086, + 0.0014077675295993686, + 0.0015092528192326427, + 0.0014332585269585252, + 0.0014244959456846118, + 0.0014626209158450365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.005249881185591221 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603143811225891 + ], + "opposite_logprob_tokens": [ + -1.9103143215179443 + ], + "logprob_reward": -0.1603143811225891, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603143811225891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06657516956329346, + 0.42938119173049927, + 0.03986218571662903, + 0.05264045670628548, + 0.07625538110733032, + 0.05954449996352196, + 0.0897490605711937, + 0.02475193701684475, + 0.13327477872371674, + 0.02796531282365322 + ], + "credit_map": [ + 0.012065638788044453, + 0.008872639387845993, + 0.006744066718965769, + 0.009197063744068146, + 0.005761928856372833, + 0.005968049168586731, + 0.006205978337675333, + 0.011759800836443901, + 0.02051786705851555, + 0.010616165585815907, + 0.1783536970615387, + 0.013198192231357098, + 0.012207324616611004, + 0.011751978658139706, + 0.006367946043610573, + 0.17636799812316895, + 0.003672567894682288, + 0.002601916203275323, + 0.010937320999801159, + 0.0031150910072028637, + 0.0019927257671952248, + 0.0025634365156292915, + 0.002729630097746849, + 0.012249497696757317, + 0.0083906389772892, + 0.007760625798255205, + 0.006240996532142162, + 0.007039584685117006, + 0.004646160639822483, + 0.00383242079988122, + 0.0046938625164330006, + 0.010036165826022625, + 0.011438010260462761, + 0.017362846061587334, + 0.006581015884876251, + 0.005847079213708639, + 0.009254854172468185, + 0.00787061545997858, + 0.008577330969274044, + 0.00932362861931324, + 0.006325965281575918, + 0.008638612926006317, + 0.008981416001915932, + 0.00981962401419878, + 0.005505233071744442, + 0.004351883661001921, + 0.007452025543898344, + 0.008469740860164165, + 0.005788424052298069, + 0.004380660131573677, + 0.028064794838428497, + 0.007157842628657818, + 0.007414404768496752, + 0.004791817627847195, + 0.005088620353490114, + 0.027062496170401573, + 0.0030441335402429104, + 0.0017800986533984542, + 0.004306437447667122, + 0.004473408218473196, + 0.0026631050277501345, + 0.002655938733369112, + 0.002256392501294613, + 0.003572422545403242, + 0.01160594541579485, + 0.00769356032833457, + 0.03578675910830498, + 0.009093261323869228, + 0.012411690317094326, + 0.01903735101222992, + 0.009516339749097824, + 0.028129873797297478, + 0.00425219489261508, + 0.005268679466098547, + 0.0036291966680437326, + 0.0024326853454113007, + 0.0036504946183413267, + 0.003339391900226474, + 0.0018958126893267035, + 0.0034968568943440914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.1603143811225891 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro close-up captured in stunning super-slow motion, executing a slow, sweeping trajectory, follows an oleander hawkmoth as it hovers with mesmerizing stability. The image reveals every minute detail of the creature's powerful wings, slowed to a hypnotic, balletic grace. The moth's torpedo-shaped body, covered in tiny, fuzzy olive-green hairs, remains almost perfectly still, while its intricate wings, frozen in time, display an intricate mosaic of olive green and khaki, cut through with sharp, geometric stripes of vibrant pink and soft white. The velvety texture of the wing scales is rendered with hyper-realistic clarity. As the camera follows the creature's subtle movements, the wings initiate a powerful downward motion, the leading edge cutting through the air and the tips twisting, revealing the complex aerodynamics at work. The background is a tapestry of dark, out-of-focus nighttime garden foliage, creating a beautiful bokeh that makes the moth's vivid colors pop. The high-speed nature photography, characterized by soft and directional lighting, captures the shimmering surface of the moth's large black eyes, casting subtle shadows that accentuate the three-dimensional forms of its body and wings, creating an atmosphere of pure curiosity focused on the mechanisms of nature.", + "seed": 18993, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_12/index_3591_seed_18993_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1231578055769205, + "vlm_logprob": -0.1231578055769205, + "vlm_yes_no_margin": 2.450000011920929, + "vlm_token_credit": -0.1231578055769205, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09741192311048508 + ], + "opposite_logprob_tokens": [ + -2.472411870956421 + ], + "logprob_reward": -0.09741192311048508, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09741192311048508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4837435483932495, + 0.07145185023546219, + 0.02391582541167736, + 0.16803842782974243, + 0.06273311376571655, + 0.024175263941287994, + 0.07314348220825195, + 0.03265712037682533, + 0.026710454374551773, + 0.033430930227041245 + ], + "credit_map": [ + 0.01287419255822897, + 0.01173388585448265, + 0.007837283425033092, + 0.00529681658372283, + 0.00984357763081789, + 0.22451935708522797, + 0.01143594365566969, + 0.20020246505737305, + 0.007687954232096672, + 0.009830518625676632, + 0.009543820284307003, + 0.012777622789144516, + 0.006960643455386162, + 0.008271103724837303, + 0.009679539129137993, + 0.006700643803924322, + 0.002467765938490629, + 0.002969790482893586, + 0.0028215739876031876, + 0.002570620970800519, + 0.003880542004480958, + 0.003164794063195586, + 0.0025962027721107006, + 0.0034445328637957573, + 0.00567217729985714, + 0.005364697892218828, + 0.01052425242960453, + 0.004964213352650404, + 0.005258878227323294, + 0.06835350394248962, + 0.0051887258887290955, + 0.06271195411682129, + 0.007567408494651318, + 0.009720990434288979, + 0.006994716357439756, + 0.010319503024220467, + 0.007139767054468393, + 0.008327700197696686, + 0.006823303643614054, + 0.005839715711772442, + 0.003111843252554536, + 0.0028008194640278816, + 0.0041636573150753975, + 0.0037523917853832245, + 0.0016761888982728124, + 0.003476636717095971, + 0.0015608238754794002, + 0.003632900770753622, + 0.006362633779644966, + 0.012298332527279854, + 0.00972015131264925, + 0.008490851148962975, + 0.006797746289521456, + 0.011236222460865974, + 0.00568100530654192, + 0.012556535191833973, + 0.004640547093003988, + 0.004613771103322506, + 0.005316728726029396, + 0.005254939664155245, + 0.0028521574568003416, + 0.004510505590587854, + 0.0020631439983844757, + 0.003405326744541526, + 0.0038860265631228685, + 0.0027045044116675854, + 0.0038270894438028336, + 0.0025968817062675953, + 0.0022776760160923004, + 0.003587468760088086, + 0.003531510941684246, + 0.004299291409552097, + 0.003816330572590232, + 0.005402399227023125, + 0.003537780838087201, + 0.004529960919171572, + 0.0027827664744108915, + 0.00226391339674592, + 0.0031045512296259403, + 0.007993222214281559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.09741192311048508 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026377661153674126 + ], + "opposite_logprob_tokens": [ + -3.6513776779174805 + ], + "logprob_reward": -0.026377661153674126, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026377661153674126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3980952799320221, + 0.12398731708526611, + 0.024112671613693237, + 0.11955899000167847, + 0.04381344094872475, + 0.035033080726861954, + 0.12736263871192932, + 0.04862941801548004, + 0.04843424633145332, + 0.030973033979535103 + ], + "credit_map": [ + 0.010293873958289623, + 0.00965105090290308, + 0.005943126976490021, + 0.0045178779400885105, + 0.007059771101921797, + 0.18538148701190948, + 0.009254558943212032, + 0.16599343717098236, + 0.01275698933750391, + 0.01525668054819107, + 0.01680382899940014, + 0.025647541508078575, + 0.012457277625799179, + 0.01370961219072342, + 0.01612049713730812, + 0.011234871111810207, + 0.0026833193842321634, + 0.002687850035727024, + 0.003130487399175763, + 0.0018639120971783996, + 0.0028810687363147736, + 0.005153293255716562, + 0.0019137614872306585, + 0.0037989732809364796, + 0.007462730631232262, + 0.011999497190117836, + 0.009039401076734066, + 0.00865362212061882, + 0.009500314481556416, + 0.033699341118335724, + 0.00933385081589222, + 0.029870210215449333, + 0.005317809991538525, + 0.007031762041151524, + 0.006330716423690319, + 0.005080148112028837, + 0.0051664067432284355, + 0.005426313728094101, + 0.004817456007003784, + 0.00464281952008605, + 0.004643961787223816, + 0.003748936578631401, + 0.004768503829836845, + 0.00503773707896471, + 0.0031560040079057217, + 0.005946318618953228, + 0.002558413427323103, + 0.005173197481781244, + 0.012189376167953014, + 0.011623588390648365, + 0.02383863925933838, + 0.0129005191847682, + 0.011135193519294262, + 0.019306061789393425, + 0.01062842458486557, + 0.025740809738636017, + 0.006031705532222986, + 0.0042998599819839, + 0.008299554698169231, + 0.005505909211933613, + 0.0038968687877058983, + 0.008533278480172157, + 0.0029385066591203213, + 0.0091237248852849, + 0.007202906999737024, + 0.005582180805504322, + 0.009053337387740612, + 0.006832507438957691, + 0.004576852545142174, + 0.004914899356663227, + 0.005505814217031002, + 0.004765735939145088, + 0.003468484617769718, + 0.0042618634179234505, + 0.005086940713226795, + 0.004840738605707884, + 0.0034645842388272285, + 0.0023012724705040455, + 0.0031093370635062456, + 0.004439808428287506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.026377661153674126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307139292359352 + ], + "opposite_logprob_tokens": [ + -3.1680715084075928 + ], + "logprob_reward": -0.04307139292359352, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307139292359352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5860636830329895, + 0.12188795208930969, + 0.018792036920785904, + 0.04575911909341812, + 0.047030217945575714, + 0.04130908474326134, + 0.039993882179260254, + 0.052570123225450516, + 0.021350380033254623, + 0.025243554264307022 + ], + "credit_map": [ + 0.015498592518270016, + 0.0142086585983634, + 0.008257327601313591, + 0.007270790636539459, + 0.010624818503856659, + 0.27258309721946716, + 0.013283684849739075, + 0.24433667957782745, + 0.01378380972892046, + 0.013590092770755291, + 0.01713157817721367, + 0.022411378100514412, + 0.012469256296753883, + 0.01432736124843359, + 0.015405804850161076, + 0.01276867650449276, + 0.0024845984298735857, + 0.003409335855394602, + 0.002214348642155528, + 0.0016002226620912552, + 0.0020779904443770647, + 0.002549591474235058, + 0.002055719029158354, + 0.0024002299178391695, + 0.0063652037642896175, + 0.008833186700940132, + 0.003969255369156599, + 0.005232584197074175, + 0.004913213662803173, + 0.006313614081591368, + 0.0040196822956204414, + 0.006112375296652317, + 0.005423099268227816, + 0.008085964247584343, + 0.0056694792583584785, + 0.0057750241830945015, + 0.005775216035544872, + 0.006537600886076689, + 0.004648277536034584, + 0.005115555599331856, + 0.0033063876908272505, + 0.0040677376091480255, + 0.006748971063643694, + 0.0037159540224820375, + 0.003193411510437727, + 0.01089765690267086, + 0.002347587374970317, + 0.007031380198895931, + 0.003551421221345663, + 0.00422505522146821, + 0.006411139387637377, + 0.004024010617285967, + 0.004240463953465223, + 0.005866175051778555, + 0.0037657227367162704, + 0.007909892126917839, + 0.007101849652826786, + 0.006229158490896225, + 0.008653035387396812, + 0.00741509348154068, + 0.0046644313260912895, + 0.00842746626585722, + 0.003173843724653125, + 0.006905247922986746, + 0.003113118465989828, + 0.0025992358569055796, + 0.002949749818071723, + 0.002565593458712101, + 0.002017596270889044, + 0.0027230840642005205, + 0.0027508677449077368, + 0.002631135517731309, + 0.0027404744178056717, + 0.003755205078050494, + 0.0036698044277727604, + 0.00352363265119493, + 0.002935031894594431, + 0.002382219536229968, + 0.002180796582251787, + 0.0040563903748989105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.04307139292359352 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34863799810409546 + ], + "opposite_logprob_tokens": [ + -1.2236380577087402 + ], + "logprob_reward": -0.34863799810409546, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34863799810409546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4823485016822815, + 0.09106703102588654, + 0.021873213350772858, + 0.08665390312671661, + 0.03414176031947136, + 0.024461574852466583, + 0.1348121613264084, + 0.031401555985212326, + 0.029424766078591347, + 0.06381549686193466 + ], + "credit_map": [ + 0.012436960823833942, + 0.011454768478870392, + 0.007732249330729246, + 0.006201985292136669, + 0.01588648557662964, + 0.2206120491027832, + 0.01196161936968565, + 0.19606244564056396, + 0.00856293085962534, + 0.010290206409990788, + 0.011316074058413506, + 0.013852539472281933, + 0.010606963187456131, + 0.011589997448027134, + 0.0112106092274189, + 0.01363772340118885, + 0.0030875899828970432, + 0.002631389070302248, + 0.0020945186261087656, + 0.001640015165321529, + 0.0026523785199970007, + 0.003944668918848038, + 0.0023524072021245956, + 0.0034702494740486145, + 0.00329745770432055, + 0.0026757842861115932, + 0.004699918907135725, + 0.0024048439227044582, + 0.002778786700218916, + 0.03571300208568573, + 0.002493730979040265, + 0.032590385526418686, + 0.003885527839884162, + 0.003690807381644845, + 0.004586408380419016, + 0.003766080131754279, + 0.0033011576160788536, + 0.004297561943531036, + 0.0042809369042515755, + 0.0063332850113511086, + 0.003435879945755005, + 0.0025360523723065853, + 0.003981008194386959, + 0.0028085708618164062, + 0.0022289715707302094, + 0.003424536669626832, + 0.0017068060114979744, + 0.004339753184467554, + 0.012815279886126518, + 0.030255228281021118, + 0.014466402120888233, + 0.020102519541978836, + 0.014494202099740505, + 0.01828777603805065, + 0.008493497036397457, + 0.015897264704108238, + 0.003769675502553582, + 0.00520658865571022, + 0.0045204502530395985, + 0.00576768396422267, + 0.002304005203768611, + 0.004436702933162451, + 0.002034819684922695, + 0.0033616353757679462, + 0.00664255628362298, + 0.004239206202328205, + 0.0042535047978162766, + 0.002157985931262374, + 0.0025627482682466507, + 0.0030714180320501328, + 0.0023567909374833107, + 0.004140559583902359, + 0.008684907108545303, + 0.009841588325798512, + 0.004730896092951298, + 0.00662975711748004, + 0.006749391555786133, + 0.00794918555766344, + 0.0051437863148748875, + 0.01408599503338337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.34863799810409546 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject an oleander hawkmoth with a torpedo-shaped body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the moth's wings colored with olive green, khaki, pink, and white stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the moth's body covered in tiny, fuzzy olive-green hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dark, out-of-focus nighttime garden foliage creating bokeh?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, sweeping trajectory following the moth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029005259275436 + ], + "opposite_logprob_tokens": [ + -2.350290060043335 + ], + "logprob_reward": -0.10029005259275436, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029005259275436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3167574405670166, + 0.06571463495492935, + 0.024894675239920616, + 0.2955675721168518, + 0.06393745541572571, + 0.05914350226521492, + 0.0413910448551178, + 0.08600186556577682, + 0.015970978885889053, + 0.030620884150266647 + ], + "credit_map": [ + 0.008842645213007927, + 0.012344283983111382, + 0.005398910958319902, + 0.005937401205301285, + 0.015618326142430305, + 0.13629122078418732, + 0.008364979177713394, + 0.12395966053009033, + 0.006013297475874424, + 0.005764144938439131, + 0.008859016001224518, + 0.013042226433753967, + 0.006786725949496031, + 0.008656611666083336, + 0.007538489066064358, + 0.009054121561348438, + 0.001897088484838605, + 0.001447291113436222, + 0.0022380435839295387, + 0.0018994236597791314, + 0.003302401164546609, + 0.005852214992046356, + 0.00196639122441411, + 0.006291819736361504, + 0.011064987629652023, + 0.01401931419968605, + 0.01911594718694687, + 0.012064128182828426, + 0.014668833464384079, + 0.10960692912340164, + 0.016635511070489883, + 0.09839191287755966, + 0.008778593502938747, + 0.012523145414888859, + 0.00752848107367754, + 0.005868646316230297, + 0.007025489583611488, + 0.008305465802550316, + 0.006891654804348946, + 0.007015973795205355, + 0.007217807695269585, + 0.007141447160393, + 0.008745010942220688, + 0.004276016727089882, + 0.005796500016003847, + 0.013665303587913513, + 0.003176697064191103, + 0.00912471767514944, + 0.005452836398035288, + 0.007845470681786537, + 0.003969899378716946, + 0.005434052553027868, + 0.00459527550265193, + 0.0060581848956644535, + 0.0028010972309857607, + 0.005234229378402233, + 0.009661605581641197, + 0.006346676032990217, + 0.015829546377062798, + 0.008150402456521988, + 0.005323953460901976, + 0.014336702413856983, + 0.0053279888816177845, + 0.02102498710155487, + 0.0024816426448524, + 0.001707028946839273, + 0.003080266062170267, + 0.0018666803371161222, + 0.0011759651824831963, + 0.002057140227407217, + 0.0012833208311349154, + 0.002318935003131628, + 0.004172796383500099, + 0.006366269662976265, + 0.0023085561115294695, + 0.0024046702310442924, + 0.0020566091407090425, + 0.002182449446991086, + 0.0018514484399929643, + 0.00927808415144682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.10029005259275436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_13/index_tensor([8639, 8639])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_13/index_tensor([8639, 8639])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a7760eb3e0846ef6a9baa0f8953ca1e1de9fc094 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_13/index_tensor([8639, 8639])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot follows a clown with an exaggerated smile as he laboriously rides a unicycle, wobbling along an empty street. The twilight glow paints the asphalt a warm, melancholic orange-gold, stretching the shadows of him and his unsteady unicycle to extreme lengths. The clown's face is thickly smeared with white paint, a large, rigid red smile stretching from the corners of his mouth to the corners of his eyes, a stark contrast to the weariness in his eyes. He wears a shaggy rainbow wig, a vibrant blue jumpsuit with giant yellow polka dots, and an exaggerated white ruffled neckpiece. To maintain his precarious balance on the swaying unicycle, he awkwardly stretches his arms, leans forward, and uses his legs with all his might to stubbornly pedal, each pedal stroke sending the unicycle lurching violently from side to side. The shutters of the shops on both sides of the street were closed, a few dead leaves were picked up by the breeze and swirled on the ground, and the bright colors of the clown formed a strong visual contrast with the monotony of the surrounding environment, creating a tranquil atmosphere that was both absurd and slightly sad.", + "seed": 30509, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_13/index_8639_seed_30509_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11141749359667301, + "vlm_logprob": -0.11141749359667301, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.11141749359667301, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20059996843338013 + ], + "opposite_logprob_tokens": [ + -2.0755999088287354 + ], + "logprob_reward": -0.20059996843338013, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.20059996843338013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14163285493850708, + 0.10411904007196426, + 0.21791259944438934, + 0.17061308026313782, + 0.043998394161462784, + 0.03235597163438797, + 0.12417533993721008, + 0.11720383912324905, + 0.03147751837968826, + 0.016511287540197372 + ], + "credit_map": [ + 0.016754623502492905, + 0.013942116871476173, + 0.020326994359493256, + 0.02112089842557907, + 0.01244360487908125, + 0.01951478235423565, + 0.025807328522205353, + 0.011722514405846596, + 0.010321171954274178, + 0.009535785764455795, + 0.014695648103952408, + 0.013765955343842506, + 0.009761789813637733, + 0.015457524918019772, + 0.01959407329559326, + 0.01098710484802723, + 0.021258117631077766, + 0.007349259685724974, + 0.06481514126062393, + 0.025481196120381355, + 0.010483955033123493, + 0.06945928186178207, + 0.006069656927138567, + 0.012996005825698376, + 0.015160989947617054, + 0.02179451286792755, + 0.017810357734560966, + 0.018275072798132896, + 0.021726394072175026, + 0.03135101869702339, + 0.022944752126932144, + 0.021549997851252556, + 0.003877890994772315, + 0.0036968793720006943, + 0.0063753859139978886, + 0.007304014638066292, + 0.0028983938973397017, + 0.005218700971454382, + 0.005023161415010691, + 0.009603971615433693, + 0.0048901657573878765, + 0.0036059743724763393, + 0.0031188682187348604, + 0.00315253552980721, + 0.0038961628451943398, + 0.0030320470687001944, + 0.007023824844509363, + 0.003636397188529372, + 0.019672393798828125, + 0.012346746399998665, + 0.01508832722902298, + 0.013481671921908855, + 0.014665779657661915, + 0.017151283100247383, + 0.011283393949270248, + 0.02048575133085251, + 0.012070821598172188, + 0.009400601498782635, + 0.015094232745468616, + 0.015363141894340515, + 0.010436590760946274, + 0.016077738255262375, + 0.011246807873249054, + 0.027513917535543442, + 0.0044846124947071075, + 0.002971835434436798, + 0.003374912543222308, + 0.003797245444729924, + 0.0049675265327095985, + 0.004372323863208294, + 0.0028299870900809765, + 0.004679078236222267, + 0.0016458934405818582, + 0.001991747412830591, + 0.0018879629205912352, + 0.0027932224329560995, + 0.002205327618867159, + 0.001666214200668037, + 0.00264534167945385, + 0.0016755786491557956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.20059996843338013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991654574871063 + ], + "opposite_logprob_tokens": [ + -1.804916501045227 + ], + "logprob_reward": -0.17991654574871063, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991654574871063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09461908787488937, + 0.10638987272977829, + 0.3882290720939636, + 0.21353919804096222, + 0.02443382889032364, + 0.030755948275327682, + 0.06088404729962349, + 0.02534106746315956, + 0.0316045880317688, + 0.02420330047607422 + ], + "credit_map": [ + 0.009117011912167072, + 0.010535961017012596, + 0.021060602739453316, + 0.01472588162869215, + 0.008082299493253231, + 0.015065109357237816, + 0.008945410139858723, + 0.007086809724569321, + 0.014439002610743046, + 0.006558877415955067, + 0.024899404495954514, + 0.014660971239209175, + 0.006515044718980789, + 0.020123416557908058, + 0.010536475107073784, + 0.008656680583953857, + 0.038258761167526245, + 0.013948315754532814, + 0.11472330242395401, + 0.045110367238521576, + 0.01651601307094097, + 0.12498172372579575, + 0.01148273516446352, + 0.023207852616906166, + 0.019132094457745552, + 0.026466215029358864, + 0.02178165130317211, + 0.02305482141673565, + 0.026130443438887596, + 0.03949647769331932, + 0.02816900983452797, + 0.02930847555398941, + 0.0027518935967236757, + 0.00244078878313303, + 0.003367953933775425, + 0.0038919327780604362, + 0.0020093906205147505, + 0.0026875666808336973, + 0.0028542194049805403, + 0.004430082626640797, + 0.004062306135892868, + 0.0030881937127560377, + 0.004049662034958601, + 0.005231494549661875, + 0.0030410615727305412, + 0.00326942908577621, + 0.0027987523935735226, + 0.005215049721300602, + 0.009698006324470043, + 0.0061605325900018215, + 0.007015157025307417, + 0.008586847223341465, + 0.007297304458916187, + 0.006898493971675634, + 0.006451117806136608, + 0.008776587434113026, + 0.002322276122868061, + 0.0019241978880017996, + 0.003335371380671859, + 0.0040372516959905624, + 0.0024824931751936674, + 0.0027063298039138317, + 0.0028120276983827353, + 0.005721120163798332, + 0.0033948959317058325, + 0.003578713396564126, + 0.0035456723999232054, + 0.0039473241195082664, + 0.005031425505876541, + 0.0052290200255811214, + 0.0030627630185335875, + 0.0038147703744471073, + 0.005210075993090868, + 0.00477116322144866, + 0.003372082719579339, + 0.003956696018576622, + 0.0015893073286861181, + 0.0012579774484038353, + 0.0016527476254850626, + 0.0023932515177875757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.17991654574871063 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043066367506980896 + ], + "opposite_logprob_tokens": [ + -3.1680662631988525 + ], + "logprob_reward": -0.043066367506980896, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043066367506980896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04482327774167061, + 0.04225848615169525, + 0.03345642238855362, + 0.6497190594673157, + 0.17484095692634583, + 0.01387563906610012, + 0.013452023267745972, + 0.009913084097206593, + 0.010165220126509666, + 0.007495894096791744 + ], + "credit_map": [ + 0.005242808721959591, + 0.007018242496997118, + 0.005548923276364803, + 0.004816464148461819, + 0.006550183519721031, + 0.005846652202308178, + 0.006504284683614969, + 0.0032957170624285936, + 0.0035918212961405516, + 0.003246285719797015, + 0.008420339785516262, + 0.0061793201602995396, + 0.004404078237712383, + 0.006681852508336306, + 0.0059999627992510796, + 0.003734821453690529, + 0.0026725081261247396, + 0.0034484751522541046, + 0.003740535117685795, + 0.0033297010231763124, + 0.0085294796153903, + 0.005801121238619089, + 0.0035205669701099396, + 0.0024140318855643272, + 0.05875929445028305, + 0.08029680699110031, + 0.06794153898954391, + 0.06950593739748001, + 0.08134502917528152, + 0.12050474435091019, + 0.08691733330488205, + 0.08444834500551224, + 0.01097834762185812, + 0.009298334829509258, + 0.03484606742858887, + 0.011822705157101154, + 0.009609072469174862, + 0.025831811130046844, + 0.010359492152929306, + 0.06209511682391167, + 0.0018193225841969252, + 0.0011609963839873672, + 0.0019317915430292487, + 0.0029217081610113382, + 0.001008850522339344, + 0.000999164185486734, + 0.0016651144251227379, + 0.002368689514696598, + 0.001420255284756422, + 0.0014770911075174809, + 0.0016273227520287037, + 0.0015592281706631184, + 0.0019580319058150053, + 0.001693113474175334, + 0.0015091936802491546, + 0.0022077863104641438, + 0.0009061865275725722, + 0.0006827773177064955, + 0.0013128452701494098, + 0.00140235701110214, + 0.0010310804937034845, + 0.0013081659562885761, + 0.0009501497261226177, + 0.002319520805031061, + 0.0010561736999079585, + 0.0011532582575455308, + 0.0012327259173616767, + 0.0015169583493843675, + 0.0015265779802575707, + 0.0012140305479988456, + 0.0011484934948384762, + 0.001317000831477344, + 0.0010160396341234446, + 0.0009174104197882116, + 0.0008715735166333616, + 0.001076027750968933, + 0.0009889970533549786, + 0.001125168171711266, + 0.0006450032815337181, + 0.0008556735701858997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.043066367506980896 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293617635965347 + ], + "opposite_logprob_tokens": [ + -2.23793625831604 + ], + "logprob_reward": -0.11293617635965347, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293617635965347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07159343361854553, + 0.08617405593395233, + 0.09608063101768494, + 0.41824713349342346, + 0.14521940052509308, + 0.01834685169160366, + 0.03396428003907204, + 0.03710993751883507, + 0.07479245960712433, + 0.018471842631697655 + ], + "credit_map": [ + 0.008740506134927273, + 0.011076127178966999, + 0.012370193377137184, + 0.010720424354076385, + 0.007001081481575966, + 0.008053477853536606, + 0.008847787976264954, + 0.004783832933753729, + 0.008778765797615051, + 0.007247801870107651, + 0.019750453531742096, + 0.012729516252875328, + 0.005607292987406254, + 0.015668034553527832, + 0.009084501303732395, + 0.007307687774300575, + 0.009162000380456448, + 0.004485536832362413, + 0.026432828977704048, + 0.011586058884859085, + 0.0069413515739142895, + 0.027703560888767242, + 0.0043173907324671745, + 0.005451899953186512, + 0.03741927072405815, + 0.05421480908989906, + 0.044405050575733185, + 0.046388160437345505, + 0.053890034556388855, + 0.07083006203174591, + 0.05591339245438576, + 0.055186353623867035, + 0.009510393254458904, + 0.008032971993088722, + 0.028899580240249634, + 0.010446067899465561, + 0.004691840149462223, + 0.022387778386473656, + 0.008531098254024982, + 0.05271966755390167, + 0.0021682034712284803, + 0.0016211834736168385, + 0.0022463174536824226, + 0.003143477952107787, + 0.002191605046391487, + 0.0019499328918755054, + 0.0025359063874930143, + 0.0024902266450226307, + 0.003444002941250801, + 0.0026295154821127653, + 0.003915107809007168, + 0.0029031059239059687, + 0.005896993447095156, + 0.005163354333490133, + 0.00412598205730319, + 0.0058862194418907166, + 0.0055225919932127, + 0.004028999712318182, + 0.004048296250402927, + 0.005568427965044975, + 0.0044431984424591064, + 0.004117014817893505, + 0.004340003710240126, + 0.005041405092924833, + 0.006596498657017946, + 0.008270050399005413, + 0.009790359996259212, + 0.007858463563024998, + 0.01500930730253458, + 0.010528613813221455, + 0.010208433493971825, + 0.006530735641717911, + 0.0024707491975277662, + 0.002031226409599185, + 0.0020321791525930166, + 0.0026104627177119255, + 0.0023720096796751022, + 0.002026690635830164, + 0.002953891409561038, + 0.0019746338948607445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.11293617635965347 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02056840993463993 + ], + "opposite_logprob_tokens": [ + -3.895568370819092 + ], + "logprob_reward": -0.02056840993463993, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02056840993463993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09622123837471008, + 0.13782986998558044, + 0.2317694127559662, + 0.19876636564731598, + 0.03939543664455414, + 0.027838127687573433, + 0.09537789225578308, + 0.050138622522354126, + 0.07162433117628098, + 0.051038727164268494 + ], + "credit_map": [ + 0.010176814161241055, + 0.010278536938130856, + 0.021260064095258713, + 0.01798797771334648, + 0.007378883194178343, + 0.014603372663259506, + 0.007740880362689495, + 0.006794711574912071, + 0.019452538341283798, + 0.00888757687062025, + 0.021121948957443237, + 0.023171182721853256, + 0.010196131654083729, + 0.030056750401854515, + 0.01265635434538126, + 0.012287386693060398, + 0.023429546505212784, + 0.007697151042521, + 0.06883696466684341, + 0.02757118083536625, + 0.009898706339299679, + 0.0742623433470726, + 0.005840905476361513, + 0.014232607558369637, + 0.017679614946246147, + 0.024044891819357872, + 0.023204559460282326, + 0.021877307444810867, + 0.024575768038630486, + 0.037621669471263885, + 0.025459861382842064, + 0.02430269494652748, + 0.005780018866062164, + 0.002814223989844322, + 0.005852547939866781, + 0.008647802285850048, + 0.0026446811389178038, + 0.0048884921707212925, + 0.00324285333044827, + 0.005524815991520882, + 0.004147983621805906, + 0.0020130188204348087, + 0.003683013143017888, + 0.0036347927525639534, + 0.0015262861270457506, + 0.0031684685964137316, + 0.0022613583132624626, + 0.00740320747718215, + 0.01208379864692688, + 0.00970240868628025, + 0.011879024095833302, + 0.011262896470725536, + 0.01023964211344719, + 0.012829194776713848, + 0.010483230464160442, + 0.016897697001695633, + 0.010259742848575115, + 0.005871869157999754, + 0.006047449540346861, + 0.009808901697397232, + 0.003255888819694519, + 0.00334992166608572, + 0.0036920292768627405, + 0.007852823473513126, + 0.012741424143314362, + 0.014339578337967396, + 0.007875231094658375, + 0.010703884065151215, + 0.007107981480658054, + 0.0062078917399048805, + 0.00471614021807909, + 0.007932201959192753, + 0.007019123062491417, + 0.01101161539554596, + 0.00712469732388854, + 0.010981063358485699, + 0.002747143851593137, + 0.002712059998884797, + 0.003168289316818118, + 0.006274733226746321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.02056840993463993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a clown with an exaggerated smile as he laboriously rides a unicycle, wobbling along an empty street. The twilight glow paints the asphalt a warm, melancholic orange-gold, stretching the shadows of him and his unsteady unicycle to extreme lengths. The clown's face is thickly smeared with white paint, a large, rigid red smile stretching from the corners of his mouth to the corners of his eyes, a stark contrast to the weariness in his eyes. He wears a shaggy rainbow wig, a vibrant blue jumpsuit with giant yellow polka dots, and an exaggerated white ruffled neckpiece. To maintain his precarious balance on the swaying unicycle, he awkwardly stretches his arms, leans forward, and uses his legs with all his might to stubbornly pedal, each pedal stroke sending the unicycle lurching violently from side to side. The shutters of the shops on both sides of the street were closed, a few dead leaves were picked up by the breeze and swirled on the ground, and the bright colors of the clown formed a strong visual contrast with the monotony of the surrounding environment, creating a tranquil atmosphere that was both absurd and slightly sad.", + "seed": 30510, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_13/index_8639_seed_30510_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1473411962389946, + "vlm_logprob": -0.1473411962389946, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.1473411962389946, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.288333922624588 + ], + "opposite_logprob_tokens": [ + -1.6633338928222656 + ], + "logprob_reward": -0.288333922624588, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.288333922624588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2586885690689087, + 0.06707324832677841, + 0.09723494201898575, + 0.0696355476975441, + 0.060831040143966675, + 0.09274069964885712, + 0.093053437769413, + 0.09564078599214554, + 0.12293344736099243, + 0.04216817766427994 + ], + "credit_map": [ + 0.02991747111082077, + 0.03354879096150398, + 0.03270052745938301, + 0.04010717570781708, + 0.030489999800920486, + 0.02897215448319912, + 0.02531728707253933, + 0.0376351960003376, + 0.009094486944377422, + 0.010264341719448566, + 0.008639374747872353, + 0.010464897379279137, + 0.008024805225431919, + 0.006026370916515589, + 0.00819283165037632, + 0.0063661495223641396, + 0.006775873247534037, + 0.01226942427456379, + 0.013646084815263748, + 0.010916980914771557, + 0.010635508224368095, + 0.024834802374243736, + 0.007599116303026676, + 0.010557161644101143, + 0.006541013717651367, + 0.008004043251276016, + 0.008521492592990398, + 0.013565406203269958, + 0.009245588444173336, + 0.008433845825493336, + 0.008554965257644653, + 0.006769197061657906, + 0.005990274250507355, + 0.006512631196528673, + 0.009461518377065659, + 0.012142723426222801, + 0.005975980311632156, + 0.00656393077224493, + 0.00694738794118166, + 0.0072365994565188885, + 0.015661437064409256, + 0.007910744287073612, + 0.012495039030909538, + 0.015170210972428322, + 0.011951141059398651, + 0.00859908014535904, + 0.01155809871852398, + 0.009394955821335316, + 0.0069785709492862225, + 0.006223313044756651, + 0.016178583726286888, + 0.011493583209812641, + 0.009599420242011547, + 0.012737789191305637, + 0.01418400276452303, + 0.0156581848859787, + 0.005454265046864748, + 0.007926827296614647, + 0.014098764397203922, + 0.020007262006402016, + 0.010007700882852077, + 0.01806189864873886, + 0.008360335603356361, + 0.01172374002635479, + 0.010422890074551105, + 0.014635112136602402, + 0.016297487542033195, + 0.014498932287096977, + 0.009630725719034672, + 0.012644276954233646, + 0.008622151799499989, + 0.03618187457323074, + 0.003949649166315794, + 0.003485222579911351, + 0.004956638906151056, + 0.005683548282831907, + 0.006823839619755745, + 0.005931726191192865, + 0.006439786870032549, + 0.004897768143564463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.288333922624588 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134687840938568 + ], + "opposite_logprob_tokens": [ + -1.3134688138961792 + ], + "logprob_reward": -0.3134687840938568, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134687840938568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22603116929531097, + 0.07660054415464401, + 0.07724843174219131, + 0.10977878421545029, + 0.09797687828540802, + 0.05477511137723923, + 0.22343914210796356, + 0.04541267454624176, + 0.04719116911292076, + 0.04154614359140396 + ], + "credit_map": [ + 0.04484542831778526, + 0.04150024428963661, + 0.02976045571267605, + 0.022659048438072205, + 0.01996622234582901, + 0.019747845828533173, + 0.015870491042733192, + 0.03168142959475517, + 0.010189149528741837, + 0.010995641350746155, + 0.009267255663871765, + 0.013948198407888412, + 0.006860118359327316, + 0.007833768613636494, + 0.00797330029308796, + 0.009533108212053776, + 0.006104317493736744, + 0.012271345593035221, + 0.010353184305131435, + 0.008604343049228191, + 0.008178873918950558, + 0.018141144886612892, + 0.004988045431673527, + 0.008607164956629276, + 0.009000460617244244, + 0.01460608933120966, + 0.01618277095258236, + 0.031164322048425674, + 0.008993582800030708, + 0.009991447441279888, + 0.009585638530552387, + 0.010254466906189919, + 0.010996272787451744, + 0.008198325522243977, + 0.014164705760776997, + 0.015021704137325287, + 0.011135539039969444, + 0.011210019700229168, + 0.008684517815709114, + 0.018565792590379715, + 0.008624928072094917, + 0.00648588128387928, + 0.006057151593267918, + 0.007953771390020847, + 0.008131093345582485, + 0.005314634181559086, + 0.005627820268273354, + 0.006579827982932329, + 0.02007531374692917, + 0.016724346205592155, + 0.0414775051176548, + 0.030682837590575218, + 0.03336934745311737, + 0.03401755541563034, + 0.023700520396232605, + 0.02339169755578041, + 0.0037560223136097193, + 0.005147965624928474, + 0.006584745831787586, + 0.008175165392458439, + 0.0051140738651156425, + 0.005749117583036423, + 0.004713313188403845, + 0.0061722672544419765, + 0.004743669182062149, + 0.00453666178509593, + 0.005928157828748226, + 0.008070982061326504, + 0.005414215847849846, + 0.006037021521478891, + 0.006046002265065908, + 0.006414451636373997, + 0.004586352035403252, + 0.004482311196625233, + 0.0053153270855546, + 0.0066345734521746635, + 0.004994876682758331, + 0.004805920645594597, + 0.005382654722779989, + 0.005344123113900423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.3134687840938568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077283769845963 + ], + "opposite_logprob_tokens": [ + -4.141077518463135 + ], + "logprob_reward": -0.016077283769845963, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077283769845963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19816461205482483, + 0.059117984026670456, + 0.09209226071834564, + 0.08218242228031158, + 0.16586244106292725, + 0.054444801062345505, + 0.15247349441051483, + 0.06994267553091049, + 0.08804253488779068, + 0.03767672926187515 + ], + "credit_map": [ + 0.03018050640821457, + 0.03138431906700134, + 0.021875962615013123, + 0.027247726917266846, + 0.02354693040251732, + 0.01843688078224659, + 0.019331231713294983, + 0.02616104669868946, + 0.005045479163527489, + 0.006412675604224205, + 0.007631531450897455, + 0.008692095056176186, + 0.009401973336935043, + 0.006739309523254633, + 0.009684388525784016, + 0.005510529037564993, + 0.007465972099453211, + 0.010098721832036972, + 0.011211656033992767, + 0.010659822262823582, + 0.021064994856715202, + 0.01364218257367611, + 0.00910936202853918, + 0.008839557878673077, + 0.00645193737000227, + 0.009933367371559143, + 0.009478878229856491, + 0.01688370108604431, + 0.013207542710006237, + 0.008557225577533245, + 0.010486015118658543, + 0.007183751091361046, + 0.009735197760164738, + 0.014228838495910168, + 0.022602463141083717, + 0.020275743678212166, + 0.039917171001434326, + 0.02913258783519268, + 0.01812419295310974, + 0.011846265755593777, + 0.005025721620768309, + 0.005637631751596928, + 0.008063158020377159, + 0.0066130505874753, + 0.00994340144097805, + 0.0039929538033902645, + 0.009988263249397278, + 0.005180619191378355, + 0.015337616205215454, + 0.011822757311165333, + 0.02721714973449707, + 0.019261177629232407, + 0.019779633730649948, + 0.016512105241417885, + 0.015314405784010887, + 0.027228649705648422, + 0.007150558289140463, + 0.011235140264034271, + 0.00861223042011261, + 0.012861888855695724, + 0.008169418200850487, + 0.008165177889168262, + 0.005920305382460356, + 0.00782795250415802, + 0.006419839337468147, + 0.007484303787350655, + 0.012388568371534348, + 0.011447290889918804, + 0.00946998130530119, + 0.015822986140847206, + 0.011201119050383568, + 0.013808456249535084, + 0.005311145912855864, + 0.004400791134685278, + 0.0046835411339998245, + 0.007130186073482037, + 0.0032948595471680164, + 0.003735333913937211, + 0.004879905376583338, + 0.004240965936332941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.016077283769845963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904305845499039 + ], + "opposite_logprob_tokens": [ + -2.464043140411377 + ], + "logprob_reward": -0.08904305845499039, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904305845499039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14835701882839203, + 0.15035580098628998, + 0.15066379308700562, + 0.10810112208127975, + 0.09562230110168457, + 0.041566066443920135, + 0.1701570600271225, + 0.057498954236507416, + 0.04531259834766388, + 0.032365236431360245 + ], + "credit_map": [ + 0.017489884048700333, + 0.026417072862386703, + 0.019909031689167023, + 0.0276559479534626, + 0.013779960572719574, + 0.012938891537487507, + 0.015629395842552185, + 0.014536838047206402, + 0.009516473859548569, + 0.017934586852788925, + 0.023617787286639214, + 0.037829913198947906, + 0.021303636953234673, + 0.00985002052038908, + 0.015650557354092598, + 0.014652830548584461, + 0.007789161521941423, + 0.013911580666899681, + 0.021173439919948578, + 0.01455723773688078, + 0.026659471914172173, + 0.038169652223587036, + 0.011119428090751171, + 0.017283814027905464, + 0.007395246531814337, + 0.013198207132518291, + 0.01698414422571659, + 0.029525427147746086, + 0.009391545318067074, + 0.009931392967700958, + 0.01312676165252924, + 0.008548389188945293, + 0.007894445210695267, + 0.010606267489492893, + 0.015449116006493568, + 0.013083706609904766, + 0.0170050710439682, + 0.012102643959224224, + 0.01149122416973114, + 0.007989831268787384, + 0.0034597846679389477, + 0.004429097753018141, + 0.005538284778594971, + 0.005440278444439173, + 0.007535399403423071, + 0.00495549151673913, + 0.005708726122975349, + 0.004499007947742939, + 0.015456211753189564, + 0.010624632239341736, + 0.03566451743245125, + 0.023142339661717415, + 0.02068191021680832, + 0.016392892226576805, + 0.018275698646903038, + 0.0299188531935215, + 0.004361655097454786, + 0.007002672180533409, + 0.008345170877873898, + 0.01465463638305664, + 0.005203542299568653, + 0.00635253731161356, + 0.005079443100839853, + 0.0064993007108569145, + 0.002551252255216241, + 0.0037144857924431562, + 0.00510751036927104, + 0.005901634693145752, + 0.008332692086696625, + 0.006005153525620699, + 0.005394499748945236, + 0.00830537173897028, + 0.0031002312898635864, + 0.0035567369777709246, + 0.004006391856819391, + 0.004460414871573448, + 0.003570443717762828, + 0.004947105422616005, + 0.0034221350215375423, + 0.005301778670400381 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.08904305845499039 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029782932251691818 + ], + "opposite_logprob_tokens": [ + -3.529783010482788 + ], + "logprob_reward": -0.029782932251691818, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029782932251691818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1290644258260727, + 0.08055241405963898, + 0.12089728564023972, + 0.15069615840911865, + 0.15855608880519867, + 0.058259665966033936, + 0.13041001558303833, + 0.04933122918009758, + 0.02701365016400814, + 0.09521905332803726 + ], + "credit_map": [ + 0.026378216221928596, + 0.021740082651376724, + 0.01513600628823042, + 0.011492484249174595, + 0.013143141753971577, + 0.013692791573703289, + 0.010791664011776447, + 0.016690045595169067, + 0.009213578887283802, + 0.009267252869904041, + 0.010344519279897213, + 0.016295120120048523, + 0.009862490929663181, + 0.006386032328009605, + 0.009225425310432911, + 0.009957997128367424, + 0.008405355736613274, + 0.0167195126414299, + 0.015720900148153305, + 0.014298065565526485, + 0.014342191629111767, + 0.02961934171617031, + 0.007080127019435167, + 0.014711805619299412, + 0.018046341836452484, + 0.02256707102060318, + 0.022994596511125565, + 0.04169085994362831, + 0.010936698876321316, + 0.011526883579790592, + 0.01130379643291235, + 0.011629922315478325, + 0.010904034599661827, + 0.023735590279102325, + 0.031997013837099075, + 0.04060746729373932, + 0.008944080211222172, + 0.009982162155210972, + 0.010718057863414288, + 0.02166769653558731, + 0.010522628203034401, + 0.0084240036085248, + 0.007958859205245972, + 0.011233448050916195, + 0.004847335163503885, + 0.00437500886619091, + 0.0040851980447769165, + 0.006813187152147293, + 0.018669355660676956, + 0.009783688932657242, + 0.02358788065612316, + 0.023189939558506012, + 0.013388477265834808, + 0.013069136999547482, + 0.012427634559571743, + 0.016293922439217567, + 0.00393054960295558, + 0.006249675992876291, + 0.0066335187293589115, + 0.010942292399704456, + 0.0061454507522284985, + 0.00607078755274415, + 0.00443617720156908, + 0.004922780208289623, + 0.0026110331527888775, + 0.003156256163492799, + 0.0032167870085686445, + 0.0033637890592217445, + 0.0036194967105984688, + 0.004241559188812971, + 0.0024423357099294662, + 0.004362394567579031, + 0.009609767235815525, + 0.012983301654458046, + 0.014743187464773655, + 0.01847812905907631, + 0.008162262849509716, + 0.012219149619340897, + 0.007893737405538559, + 0.011129525490105152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.029782932251691818 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a clown with an exaggerated smile as he laboriously rides a unicycle, wobbling along an empty street. The twilight glow paints the asphalt a warm, melancholic orange-gold, stretching the shadows of him and his unsteady unicycle to extreme lengths. The clown's face is thickly smeared with white paint, a large, rigid red smile stretching from the corners of his mouth to the corners of his eyes, a stark contrast to the weariness in his eyes. He wears a shaggy rainbow wig, a vibrant blue jumpsuit with giant yellow polka dots, and an exaggerated white ruffled neckpiece. To maintain his precarious balance on the swaying unicycle, he awkwardly stretches his arms, leans forward, and uses his legs with all his might to stubbornly pedal, each pedal stroke sending the unicycle lurching violently from side to side. The shutters of the shops on both sides of the street were closed, a few dead leaves were picked up by the breeze and swirled on the ground, and the bright colors of the clown formed a strong visual contrast with the monotony of the surrounding environment, creating a tranquil atmosphere that was both absurd and slightly sad.", + "seed": 30511, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_13/index_8639_seed_30511_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16001031808555127, + "vlm_logprob": -0.16001031808555127, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.16001031808555127, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25716134905815125 + ], + "opposite_logprob_tokens": [ + -1.7571613788604736 + ], + "logprob_reward": -0.25716134905815125, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.25716134905815125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12545113265514374, + 0.031168971210718155, + 0.06168605759739876, + 0.11357754468917847, + 0.08803701400756836, + 0.10549930483102798, + 0.10363016277551651, + 0.171214759349823, + 0.12769076228141785, + 0.0720442607998848 + ], + "credit_map": [ + 0.01596466451883316, + 0.008146875537931919, + 0.025398189201951027, + 0.017471984028816223, + 0.011359335854649544, + 0.013074689544737339, + 0.017692934721708298, + 0.01634245552122593, + 0.005413735285401344, + 0.002514269668608904, + 0.004851122386753559, + 0.0044492920860648155, + 0.0027692581061273813, + 0.0028336653485894203, + 0.004616898484528065, + 0.003720731008797884, + 0.006084836088120937, + 0.01018141582608223, + 0.009011367335915565, + 0.009229565970599651, + 0.008186300285160542, + 0.004422985482960939, + 0.009523678570985794, + 0.005045906640589237, + 0.016187842935323715, + 0.020567385479807854, + 0.012822389602661133, + 0.013428620994091034, + 0.012884126976132393, + 0.010946724563837051, + 0.011969462968409061, + 0.01477098185569048, + 0.006101015489548445, + 0.004975827876478434, + 0.009251316078007221, + 0.02428482472896576, + 0.006006456445902586, + 0.022154251113533974, + 0.007019958924502134, + 0.008243358694016933, + 0.006817162502557039, + 0.005794618744403124, + 0.010966003872454166, + 0.029665354639291763, + 0.0045626359060406685, + 0.032678890973329544, + 0.007347149774432182, + 0.0076674893498420715, + 0.015076062642037868, + 0.010456087067723274, + 0.02002941258251667, + 0.01613231748342514, + 0.008244089782238007, + 0.012709897942841053, + 0.011113365180790424, + 0.009868928231298923, + 0.02519373781979084, + 0.019026752561330795, + 0.020794587209820747, + 0.02334410510957241, + 0.01661030575633049, + 0.026097575202584267, + 0.024021493270993233, + 0.016126208007335663, + 0.011075431481003761, + 0.018677351996302605, + 0.01626385748386383, + 0.029869310557842255, + 0.015712497755885124, + 0.008897327817976475, + 0.013419189490377903, + 0.01377579104155302, + 0.007301004137843847, + 0.010017039254307747, + 0.010748747736215591, + 0.008961389772593975, + 0.006732537876814604, + 0.009340351447463036, + 0.008716032840311527, + 0.010227156803011894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.25716134905815125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486688435077667 + ], + "opposite_logprob_tokens": [ + -1.2236688137054443 + ], + "logprob_reward": -0.3486688435077667, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486688435077667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13486015796661377, + 0.03584900125861168, + 0.08967479318380356, + 0.14504101872444153, + 0.0942763090133667, + 0.08307256549596786, + 0.07826121151447296, + 0.22625601291656494, + 0.079018734395504, + 0.03369024023413658 + ], + "credit_map": [ + 0.01400802843272686, + 0.021213117986917496, + 0.02179049327969551, + 0.016532234847545624, + 0.012070739641785622, + 0.01417019683867693, + 0.01956881396472454, + 0.01550651341676712, + 0.005876501090824604, + 0.004068004433065653, + 0.0047377352602779865, + 0.003982441499829292, + 0.0047266255132853985, + 0.0032930118031799793, + 0.005705839488655329, + 0.003458840074017644, + 0.01139698177576065, + 0.012377419508993626, + 0.015631090849637985, + 0.011110215447843075, + 0.013350142166018486, + 0.006510202772915363, + 0.010436167009174824, + 0.008862567134201527, + 0.0171677153557539, + 0.023504499346017838, + 0.019804690033197403, + 0.020976396277546883, + 0.018382256850600243, + 0.012934789061546326, + 0.017889024689793587, + 0.014381635002791882, + 0.0066875736229121685, + 0.00605096947401762, + 0.016889261081814766, + 0.018043791875243187, + 0.0072692385874688625, + 0.021375110372900963, + 0.006466328632086515, + 0.011494026519358158, + 0.005769902374595404, + 0.005944954697042704, + 0.009089654311537743, + 0.022964389994740486, + 0.0038544752169400454, + 0.024759043008089066, + 0.004708795342594385, + 0.0059813400730490685, + 0.010241638869047165, + 0.010098681785166264, + 0.015349628403782845, + 0.011412401683628559, + 0.005767911672592163, + 0.008085101842880249, + 0.011264634318649769, + 0.006041205488145351, + 0.02253318391740322, + 0.027284856885671616, + 0.026935884729027748, + 0.03572839871048927, + 0.025500711053609848, + 0.030823171138763428, + 0.03721160441637039, + 0.02023817040026188, + 0.007247540634125471, + 0.01175173930823803, + 0.011125536635518074, + 0.014807034283876419, + 0.006656909827142954, + 0.0058443378657102585, + 0.00968479085713625, + 0.011900837533175945, + 0.0035668332129716873, + 0.004069468937814236, + 0.004859340377151966, + 0.0037149691488593817, + 0.0032749143429100513, + 0.0060522411949932575, + 0.0035199110861867666, + 0.004632559139281511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.3486688435077667 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350026950240135 + ], + "opposite_logprob_tokens": [ + -3.6513500213623047 + ], + "logprob_reward": -0.026350026950240135, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350026950240135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0917162373661995, + 0.02896200492978096, + 0.04659479856491089, + 0.14854983985424042, + 0.1369188278913498, + 0.34615379571914673, + 0.0481996014714241, + 0.025724204257130623, + 0.056879010051488876, + 0.07030165940523148 + ], + "credit_map": [ + 0.013548568822443485, + 0.009796463884413242, + 0.010739438235759735, + 0.011149079538881779, + 0.00905772764235735, + 0.012308930978178978, + 0.013473080471158028, + 0.011642960831522942, + 0.0034413172397762537, + 0.0022122180089354515, + 0.005906966514885426, + 0.0038875204045325518, + 0.001849469030275941, + 0.003025723621249199, + 0.0038946429267525673, + 0.004744148813188076, + 0.004290974698960781, + 0.007763446308672428, + 0.006750298198312521, + 0.009009424597024918, + 0.004906959366053343, + 0.0035474810283631086, + 0.006789248902350664, + 0.0035369661636650562, + 0.023615922778844833, + 0.022374091669917107, + 0.024547934532165527, + 0.013941614888608456, + 0.013448111712932587, + 0.0199775081127882, + 0.012556737288832664, + 0.018087929114699364, + 0.007207756862044334, + 0.007193945813924074, + 0.03156376630067825, + 0.04790433496236801, + 0.00804749596863985, + 0.008511297404766083, + 0.007519716862589121, + 0.018970534205436707, + 0.024885527789592743, + 0.01644199527800083, + 0.03388945385813713, + 0.10118957608938217, + 0.01909502036869526, + 0.10208657383918762, + 0.025278545916080475, + 0.023287106305360794, + 0.004852360114455223, + 0.0069606793113052845, + 0.007870475761592388, + 0.008345483802258968, + 0.005699419416487217, + 0.004539916291832924, + 0.006075836718082428, + 0.0038554328493773937, + 0.002963088685646653, + 0.002275176113471389, + 0.004105031490325928, + 0.004110747016966343, + 0.0016437281155958772, + 0.0020315167494118214, + 0.006006425246596336, + 0.002588490257039666, + 0.004206527024507523, + 0.007653185166418552, + 0.00853313971310854, + 0.010508446954190731, + 0.005156554281711578, + 0.005167451221495867, + 0.00824864860624075, + 0.007405064068734646, + 0.006063597276806831, + 0.01341584324836731, + 0.007697825785726309, + 0.012975762598216534, + 0.008871800266206264, + 0.00536143034696579, + 0.00873297918587923, + 0.007182425819337368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.026350026950240135 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495350807905197 + ], + "opposite_logprob_tokens": [ + -2.9299535751342773 + ], + "logprob_reward": -0.05495350807905197, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495350807905197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1931663453578949, + 0.04612703248858452, + 0.12923461198806763, + 0.0912967100739479, + 0.05076664313673973, + 0.13980595767498016, + 0.07272128760814667, + 0.0771482065320015, + 0.13037921488285065, + 0.06935402750968933 + ], + "credit_map": [ + 0.022653643041849136, + 0.01210001390427351, + 0.03208721801638603, + 0.024800127372145653, + 0.01825088821351528, + 0.015573263168334961, + 0.04843238741159439, + 0.01926877535879612, + 0.005475272424519062, + 0.0032460843212902546, + 0.006970536429435015, + 0.006544416304677725, + 0.004567150492221117, + 0.0038353914860635996, + 0.008600280620157719, + 0.00688789552077651, + 0.012391402386128902, + 0.01938812807202339, + 0.01792350597679615, + 0.019210468977689743, + 0.01330089196562767, + 0.0066655767150223255, + 0.02970513142645359, + 0.010649500414729118, + 0.013055876828730106, + 0.016468096524477005, + 0.013797169551253319, + 0.0097516980022192, + 0.009074656292796135, + 0.010410880669951439, + 0.0077688200399279594, + 0.010969499126076698, + 0.0042749145068228245, + 0.00308238691650331, + 0.008966593071818352, + 0.0036771188024431467, + 0.004015402402728796, + 0.015407534316182137, + 0.004402081482112408, + 0.006940606981515884, + 0.009859632700681686, + 0.0072576687671244144, + 0.015163629315793514, + 0.040164269506931305, + 0.0061555663123726845, + 0.04266471415758133, + 0.008278408087790012, + 0.01026206649839878, + 0.007271215785294771, + 0.007885902188718319, + 0.014790338464081287, + 0.014307256788015366, + 0.006125333718955517, + 0.006229201331734657, + 0.010120037011802197, + 0.005991996731609106, + 0.010560027323663235, + 0.0068698241375386715, + 0.012383533641695976, + 0.012895136140286922, + 0.006389251444488764, + 0.00668740039691329, + 0.01453323569148779, + 0.006829787511378527, + 0.011171265505254269, + 0.01976778730750084, + 0.016940657049417496, + 0.030161671340465546, + 0.013902703300118446, + 0.009794006124138832, + 0.01696661487221718, + 0.011674496345221996, + 0.008057188242673874, + 0.008186583407223225, + 0.009743095375597477, + 0.010465533472597599, + 0.006817736197263002, + 0.005717874504625797, + 0.010757802985608578, + 0.009608202613890171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.05495350807905197 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291786283254623 + ], + "opposite_logprob_tokens": [ + -2.237917900085449 + ], + "logprob_reward": -0.11291786283254623, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291786283254623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14565664529800415, + 0.052515264600515366, + 0.041930630803108215, + 0.196116641163826, + 0.06083983927965164, + 0.1355774849653244, + 0.14965230226516724, + 0.04624628275632858, + 0.07903936505317688, + 0.09242548793554306 + ], + "credit_map": [ + 0.024211784824728966, + 0.016829567030072212, + 0.02001372165977955, + 0.015954159200191498, + 0.012897486798465252, + 0.022569317370653152, + 0.014057782478630543, + 0.019122840836644173, + 0.007511952891945839, + 0.0035191457718610764, + 0.010461803525686264, + 0.0066696410067379475, + 0.003831788431853056, + 0.005896293092519045, + 0.005863531958311796, + 0.008761110715568066, + 0.005279623903334141, + 0.005813256371766329, + 0.0063051325269043446, + 0.00600755400955677, + 0.005739845335483551, + 0.00314413127489388, + 0.005644797347486019, + 0.003996295854449272, + 0.028679106384515762, + 0.035917673259973526, + 0.026002736762166023, + 0.02100892923772335, + 0.020491357892751694, + 0.020954465493559837, + 0.01846821792423725, + 0.02459414303302765, + 0.003470195457339287, + 0.005379725247621536, + 0.012287952937185764, + 0.020729120820760727, + 0.003360886126756668, + 0.004807912278920412, + 0.003513711504638195, + 0.007290340960025787, + 0.01473390031605959, + 0.00653032585978508, + 0.016660384833812714, + 0.03598790988326073, + 0.005652386695146561, + 0.0388261154294014, + 0.007297010160982609, + 0.009889461100101471, + 0.02243739925324917, + 0.01460591796785593, + 0.029107563197612762, + 0.02251517027616501, + 0.010410910472273827, + 0.01749519817531109, + 0.01831548847258091, + 0.014764664694666862, + 0.005073182284832001, + 0.006034873425960541, + 0.006794341374188662, + 0.006811007857322693, + 0.0036629021633416414, + 0.006563317961990833, + 0.006230376660823822, + 0.005076282192021608, + 0.00804227776825428, + 0.0062663014978170395, + 0.011863330379128456, + 0.009886328130960464, + 0.004600973334163427, + 0.0122829619795084, + 0.010042521171271801, + 0.016054674983024597, + 0.010597456246614456, + 0.007693757303059101, + 0.018383361399173737, + 0.010272607207298279, + 0.005776525475084782, + 0.013363410718739033, + 0.009864619933068752, + 0.016473757103085518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.11291786283254623 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a clown with an exaggerated smile as he laboriously rides a unicycle, wobbling along an empty street. The twilight glow paints the asphalt a warm, melancholic orange-gold, stretching the shadows of him and his unsteady unicycle to extreme lengths. The clown's face is thickly smeared with white paint, a large, rigid red smile stretching from the corners of his mouth to the corners of his eyes, a stark contrast to the weariness in his eyes. He wears a shaggy rainbow wig, a vibrant blue jumpsuit with giant yellow polka dots, and an exaggerated white ruffled neckpiece. To maintain his precarious balance on the swaying unicycle, he awkwardly stretches his arms, leans forward, and uses his legs with all his might to stubbornly pedal, each pedal stroke sending the unicycle lurching violently from side to side. The shutters of the shops on both sides of the street were closed, a few dead leaves were picked up by the breeze and swirled on the ground, and the bright colors of the clown formed a strong visual contrast with the monotony of the surrounding environment, creating a tranquil atmosphere that was both absurd and slightly sad.", + "seed": 30512, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_13/index_8639_seed_30512_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15870949439704418, + "vlm_logprob": -0.15870949439704418, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.15870949439704418, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19233199954032898 + ], + "opposite_logprob_tokens": [ + -2.1923320293426514 + ], + "logprob_reward": -0.19233199954032898, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.19233199954032898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22767388820648193, + 0.05404244363307953, + 0.08592286705970764, + 0.0352741964161396, + 0.05422468110918999, + 0.02239375002682209, + 0.0304095596075058, + 0.06339526921510696, + 0.3594864308834076, + 0.06717681884765625 + ], + "credit_map": [ + 0.013790316879749298, + 0.020930442959070206, + 0.03988407179713249, + 0.034398846328258514, + 0.020288744941353798, + 0.05574074387550354, + 0.022334938868880272, + 0.020305806770920753, + 0.005787061993032694, + 0.0077368346974253654, + 0.007972770370543003, + 0.008401969447731972, + 0.0052458541467785835, + 0.005466631148010492, + 0.007619533687829971, + 0.005811798386275768, + 0.012232674285769463, + 0.00794215314090252, + 0.014886711724102497, + 0.012681691907346249, + 0.007790048606693745, + 0.01291689369827509, + 0.00833315122872591, + 0.00913955271244049, + 0.005949449725449085, + 0.003993320744484663, + 0.005484640132635832, + 0.005947335157543421, + 0.0024720055516809225, + 0.004223459865897894, + 0.003107707714661956, + 0.004096280783414841, + 0.00414296006783843, + 0.005909888539463282, + 0.008075702004134655, + 0.007896238006651402, + 0.008500420488417149, + 0.006145659368485212, + 0.006938244216144085, + 0.00661557586863637, + 0.0026859284844249487, + 0.002124920254573226, + 0.0036976186092942953, + 0.004183124750852585, + 0.001837690593674779, + 0.002897819736972451, + 0.0015888531925156713, + 0.003377797082066536, + 0.003620984498411417, + 0.004297792911529541, + 0.004662622231990099, + 0.004568157717585564, + 0.003218358615413308, + 0.0034536367747932673, + 0.003267447231337428, + 0.0033205626532435417, + 0.006919024046510458, + 0.011070005595684052, + 0.010476860217750072, + 0.012348691001534462, + 0.006777047645300627, + 0.003977437969297171, + 0.005122659727931023, + 0.006703550461679697, + 0.03831568360328674, + 0.05252007395029068, + 0.05131237953901291, + 0.06493521481752396, + 0.039617475122213364, + 0.030968591570854187, + 0.03260552138090134, + 0.0492115318775177, + 0.006843749899417162, + 0.012439708225429058, + 0.0105652566999197, + 0.007465507369488478, + 0.007183847017586231, + 0.009845455177128315, + 0.005292050540447235, + 0.007541251834481955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.19233199954032898 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289276599884033 + ], + "opposite_logprob_tokens": [ + -1.0539276599884033 + ], + "logprob_reward": -0.4289276599884033, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289276599884033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23494946956634521, + 0.09838370978832245, + 0.07634951174259186, + 0.030542422086000443, + 0.04040592536330223, + 0.02647358737885952, + 0.03286559134721756, + 0.03849629685282707, + 0.11404683440923691, + 0.3074866533279419 + ], + "credit_map": [ + 0.017514336854219437, + 0.0276032704859972, + 0.031819898635149, + 0.0290692038834095, + 0.024652542546391487, + 0.01888594962656498, + 0.053844768553972244, + 0.03155951946973801, + 0.008361601270735264, + 0.015149635262787342, + 0.01552487164735794, + 0.012507282197475433, + 0.015847351402044296, + 0.007664681412279606, + 0.013391493819653988, + 0.009936797432601452, + 0.006725048180669546, + 0.005488889757543802, + 0.012189467437565327, + 0.00813991203904152, + 0.010979801416397095, + 0.013440347276628017, + 0.011380276642739773, + 0.008005773648619652, + 0.0023461952805519104, + 0.0034021944738924503, + 0.0054008993320167065, + 0.005998692475259304, + 0.002190573140978813, + 0.003111479338258505, + 0.004049669951200485, + 0.0040427204221487045, + 0.0028816997073590755, + 0.0036095071118324995, + 0.005693242885172367, + 0.006174350157380104, + 0.005579897202551365, + 0.005155876744538546, + 0.005118479486554861, + 0.006192873697727919, + 0.003358283545821905, + 0.0029036439955234528, + 0.003290710737928748, + 0.0033209335524588823, + 0.0030351513996720314, + 0.0030023399740457535, + 0.0039550443179905415, + 0.0036074828822165728, + 0.0028577742632478476, + 0.002392154186964035, + 0.005199886858463287, + 0.00611433619633317, + 0.003255035961046815, + 0.006009915377944708, + 0.0042479513213038445, + 0.0027885399758815765, + 0.003241842845454812, + 0.005266204476356506, + 0.005144672468304634, + 0.005121896974742413, + 0.0049066380597651005, + 0.0030997367575764656, + 0.00712891248986125, + 0.004586395807564259, + 0.015163534320890903, + 0.01947087049484253, + 0.015224475413560867, + 0.013167756609618664, + 0.01265516597777605, + 0.011082719080150127, + 0.01683419942855835, + 0.010448123328387737, + 0.026480091735720634, + 0.029113803058862686, + 0.04264862462878227, + 0.03311358764767647, + 0.035659823566675186, + 0.07463125139474869, + 0.030522599816322327, + 0.035316865891218185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.4289276599884033 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297740146517754 + ], + "opposite_logprob_tokens": [ + -3.7732977867126465 + ], + "logprob_reward": -0.023297740146517754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297740146517754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37290865182876587, + 0.12271561473608017, + 0.0837530866265297, + 0.1003347858786583, + 0.07488290965557098, + 0.02712702751159668, + 0.028202496469020844, + 0.021448764950037003, + 0.048657819628715515, + 0.11996884644031525 + ], + "credit_map": [ + 0.020279495045542717, + 0.03459928184747696, + 0.0570976622402668, + 0.06155339255928993, + 0.029580550268292427, + 0.028825899586081505, + 0.09354358911514282, + 0.04742879420518875, + 0.012052603997290134, + 0.00834186002612114, + 0.017600469291210175, + 0.01333000510931015, + 0.014518769457936287, + 0.019067419692873955, + 0.024136971682310104, + 0.01366751454770565, + 0.009398423135280609, + 0.004661113489419222, + 0.01306298840790987, + 0.012506529688835144, + 0.011454237625002861, + 0.012334226630628109, + 0.012442836537957191, + 0.00789274275302887, + 0.005050138104707003, + 0.004762547090649605, + 0.01656046323478222, + 0.016818763688206673, + 0.015187890268862247, + 0.010121133178472519, + 0.02160761132836342, + 0.01022624596953392, + 0.007808083668351173, + 0.008233298547565937, + 0.010176202282309532, + 0.00880346354097128, + 0.009799384512007236, + 0.007486927323043346, + 0.0131285535171628, + 0.009447001852095127, + 0.001938515342772007, + 0.0033668489195406437, + 0.0038237939588725567, + 0.0037178003694862127, + 0.004415118135511875, + 0.0029099059756845236, + 0.003960058558732271, + 0.0029949876479804516, + 0.004454575013369322, + 0.0027453992515802383, + 0.003779186401516199, + 0.0049155354499816895, + 0.0031827804632484913, + 0.0027915439568459988, + 0.0031456430442631245, + 0.0031878340523689985, + 0.0022555775940418243, + 0.00273763295263052, + 0.0028961091302335262, + 0.002802459988743067, + 0.0028644234407693148, + 0.001721287495456636, + 0.0038228544872254133, + 0.0023484209086745977, + 0.006725228391587734, + 0.006424658000469208, + 0.006850326433777809, + 0.005911845713853836, + 0.006335692945867777, + 0.005136858206242323, + 0.006974620278924704, + 0.004298596177250147, + 0.012921037152409554, + 0.015442014671862125, + 0.017331384122371674, + 0.016589615494012833, + 0.011903227306902409, + 0.02199033834040165, + 0.010365002788603306, + 0.013426227495074272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.023297740146517754 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002981752157211 + ], + "opposite_logprob_tokens": [ + -2.6950297355651855 + ], + "logprob_reward": -0.07002981752157211, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002981752157211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2852450907230377, + 0.09401023387908936, + 0.07784950733184814, + 0.06135115399956703, + 0.05696622282266617, + 0.05637577548623085, + 0.045009322464466095, + 0.031467124819755554, + 0.0737614706158638, + 0.21796415746212006 + ], + "credit_map": [ + 0.018059078603982925, + 0.01882559433579445, + 0.042764876037836075, + 0.042214781045913696, + 0.04066043719649315, + 0.03164270147681236, + 0.059024177491664886, + 0.03205343335866928, + 0.00770863750949502, + 0.01175810769200325, + 0.014005879871547222, + 0.012168300338089466, + 0.01227083895355463, + 0.008953445591032505, + 0.01818392425775528, + 0.008961097337305546, + 0.00737248407676816, + 0.006222110707312822, + 0.012107037007808685, + 0.007055160589516163, + 0.009596338495612144, + 0.016117442399263382, + 0.009149083867669106, + 0.01022984180599451, + 0.0046965591609478, + 0.0049828654155135155, + 0.008773252367973328, + 0.010342258028686047, + 0.008939208462834358, + 0.006048197392374277, + 0.012254939414560795, + 0.0053138695657253265, + 0.00603589927777648, + 0.007590451277792454, + 0.008138414472341537, + 0.0076391627080738544, + 0.007812684401869774, + 0.004738372750580311, + 0.009081781841814518, + 0.005929454695433378, + 0.008527673780918121, + 0.004696951713413, + 0.008423234336078167, + 0.005852234084159136, + 0.006195052992552519, + 0.006701345555484295, + 0.007707662880420685, + 0.008271612226963043, + 0.003841218538582325, + 0.004275960382074118, + 0.006551410537213087, + 0.0074117109179496765, + 0.004914932884275913, + 0.007359574548900127, + 0.005385882221162319, + 0.005268627777695656, + 0.004806592594832182, + 0.003965368028730154, + 0.004666103981435299, + 0.004343618638813496, + 0.0034245962742716074, + 0.002194230677559972, + 0.004624971654266119, + 0.0034416394773870707, + 0.009503858163952827, + 0.012490362860262394, + 0.00988767296075821, + 0.010505205020308495, + 0.008845509961247444, + 0.006885855458676815, + 0.0090207289904356, + 0.006622267421334982, + 0.017151683568954468, + 0.01879667304456234, + 0.02709362469613552, + 0.023864980787038803, + 0.04208415001630783, + 0.023961208760738373, + 0.0363851860165596, + 0.02862662822008133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07002981752157211 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following the clown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the clown wearing a shaggy rainbow wig and a blue jumpsuit with giant yellow polka dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a unicycle while wobbling and lurching violently from side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment an empty street with closed shop shutters and a twilight glow painting the asphalt orange-gold?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896025478839874 + ], + "opposite_logprob_tokens": [ + -2.5789601802825928 + ], + "logprob_reward": -0.07896025478839874, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896025478839874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2284935861825943, + 0.03689376637339592, + 0.10603547841310501, + 0.050284840166568756, + 0.08703426271677017, + 0.07205498963594437, + 0.03209036588668823, + 0.03874551132321358, + 0.046765025705099106, + 0.30160221457481384 + ], + "credit_map": [ + 0.016297750174999237, + 0.021432461217045784, + 0.04250895231962204, + 0.055156875401735306, + 0.018608348444104195, + 0.02572571486234665, + 0.030411213636398315, + 0.018352264538407326, + 0.0044411891140043736, + 0.004080609884113073, + 0.00570104643702507, + 0.005187489558011293, + 0.0037097784224897623, + 0.004067853093147278, + 0.005417877808213234, + 0.004287919495254755, + 0.016003374010324478, + 0.01165362261235714, + 0.017801646143198013, + 0.02279655635356903, + 0.008161710575222969, + 0.009945515543222427, + 0.009939978830516338, + 0.009733077138662338, + 0.008962297812104225, + 0.004352150019258261, + 0.007825826294720173, + 0.008259064517915249, + 0.004695716314017773, + 0.006763854529708624, + 0.003900351468473673, + 0.0055255768820643425, + 0.012184266932308674, + 0.00935672502964735, + 0.012798100709915161, + 0.0161240603774786, + 0.008572281338274479, + 0.011239001527428627, + 0.00714582484215498, + 0.009613992646336555, + 0.013892979361116886, + 0.011763320304453373, + 0.010745595209300518, + 0.009815615601837635, + 0.005707095842808485, + 0.007290820591151714, + 0.005053489934653044, + 0.0077860718593001366, + 0.005624394863843918, + 0.004793152213096619, + 0.0044025820679962635, + 0.005279368255287409, + 0.0028501187916845083, + 0.0037518483586609364, + 0.0030885834712535143, + 0.0023003185633569956, + 0.00409423653036356, + 0.006325596943497658, + 0.0063012707978487015, + 0.006334770005196333, + 0.004436499904841185, + 0.0027900056447833776, + 0.0034189410507678986, + 0.005044188350439072, + 0.0054579260759055614, + 0.0070937760174274445, + 0.006887039635330439, + 0.0067749167792499065, + 0.004706724546849728, + 0.005437676794826984, + 0.0049450290389359, + 0.005461940076202154, + 0.0364677794277668, + 0.051899950951337814, + 0.05091216415166855, + 0.04399620741605759, + 0.022885115817189217, + 0.028182486072182655, + 0.01920095644891262, + 0.04805755615234375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.07896025478839874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_14/index_tensor([6099, 6099])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_14/index_tensor([6099, 6099])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..887668914484693cdc45677fdd474b8a18e685fd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_14/index_tensor([6099, 6099])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in cinematic extreme close-up with a macro lens, the film features a slow and graceful 180-degree orbiting shot around a single candle flame. The candle, a simple column of white wax with a subtle dry-drip texture on its sides, rests on a dark, reflective wooden surface within a pitch-black room. Initially, its blackened wick supports only a feeble, sputtering flame—a tiny, almost translucent teardrop of pale yellow and blue light that flickers hesitantly, casting faint, trembling shadows. As the camera steadily orbits the subject, the flame undergoes a sudden, dramatic transformation; in a swift, decisive moment, it \"wakes up,\" instantly reaching upward and intensifying into a steady, confident, almond-shaped flame of brilliant orange-yellow, with a vibrant blue core at its base. This newly awakened flame now stands firm, radiating a powerful, warm, golden glow that pools on the candle's waxy surface and casts a soft, steady glow on the polished wood beneath, creating an atmosphere of deep intimacy and contemplation, completely focused on this singular moment of burning and its tranquil glow.", + "seed": 15263, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_14/index_6099_seed_15263_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.5207902073860169, + "vlm_logprob": -0.5207902073860169, + "vlm_yes_no_margin": 1.575, + "vlm_token_credit": -0.5207902073860169, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03612286224961281 + ], + "opposite_logprob_tokens": [ + -3.6611227989196777 + ], + "logprob_reward": -0.03612286224961281, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03612286224961281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19158710539340973, + 0.04590647667646408, + 0.10567901283502579, + 0.07992995530366898, + 0.27669328451156616, + 0.10695996880531311, + 0.04983455315232277, + 0.08952230960130692, + 0.03032948076725006, + 0.023557810112833977 + ], + "credit_map": [ + 0.017919989302754402, + 0.021443011239171028, + 0.02005464769899845, + 0.02557065151631832, + 0.012490262277424335, + 0.03753391653299332, + 0.0180776659399271, + 0.038496967405080795, + 0.00413088034838438, + 0.004792565479874611, + 0.004260166082531214, + 0.004390675108879805, + 0.0032483753748238087, + 0.009987425990402699, + 0.003958724904805422, + 0.011137660592794418, + 0.009294227696955204, + 0.011662409640848637, + 0.008301452733576298, + 0.00800099316984415, + 0.0073398323729634285, + 0.025028537958860397, + 0.00789127591997385, + 0.028160283342003822, + 0.008205520920455456, + 0.004565167706459761, + 0.006610031705349684, + 0.006176668684929609, + 0.004288061056286097, + 0.018091076985001564, + 0.005632929503917694, + 0.026360489428043365, + 0.028660712763667107, + 0.0225386843085289, + 0.021365806460380554, + 0.026654748246073723, + 0.018493304029107094, + 0.047636374831199646, + 0.034230995923280716, + 0.07711264491081238, + 0.010730491019785404, + 0.004916712176054716, + 0.008764000609517097, + 0.009102463722229004, + 0.007315327413380146, + 0.021758059039711952, + 0.01580844633281231, + 0.028564469888806343, + 0.006211151368916035, + 0.0029208443593233824, + 0.00881007220596075, + 0.004152955953031778, + 0.0052988179959356785, + 0.007226940710097551, + 0.006642742082476616, + 0.008571028709411621, + 0.017332907766103745, + 0.003018229501321912, + 0.00855285208672285, + 0.00597618892788887, + 0.008872791193425655, + 0.018443774431943893, + 0.006264574360102415, + 0.021060990169644356, + 0.00480685755610466, + 0.0013554678298532963, + 0.002737591043114662, + 0.0026065807323902845, + 0.00409464119002223, + 0.004189090803265572, + 0.0033082175068557262, + 0.007231033407151699, + 0.002923219930380583, + 0.001055148895829916, + 0.0025074761360883713, + 0.0016244676662608981, + 0.002268934855237603, + 0.0039282809011638165, + 0.0022410068195313215, + 0.007009273394942284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.03612286224961281 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372875973582268 + ], + "opposite_logprob_tokens": [ + -3.408728837966919 + ], + "logprob_reward": -0.03372875973582268, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372875973582268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15310806035995483, + 0.042992811650037766, + 0.1323876976966858, + 0.07720249146223068, + 0.19941389560699463, + 0.07004407048225403, + 0.11184826493263245, + 0.10640247166156769, + 0.06295596063137054, + 0.0436442494392395 + ], + "credit_map": [ + 0.014251782558858395, + 0.0140896150842309, + 0.02205769717693329, + 0.022809382528066635, + 0.00887299794703722, + 0.026312943547964096, + 0.0133096007630229, + 0.0314040407538414, + 0.0032749930396676064, + 0.005449164658784866, + 0.003147399052977562, + 0.004298849496990442, + 0.003185501554980874, + 0.008752169087529182, + 0.004008864052593708, + 0.010875874198973179, + 0.010274460539221764, + 0.011454620398581028, + 0.01795016974210739, + 0.010892726480960846, + 0.011231163516640663, + 0.024749092757701874, + 0.014385486021637917, + 0.03144996985793114, + 0.00813219789415598, + 0.006494014989584684, + 0.008076304569840431, + 0.007432479877024889, + 0.005109372083097696, + 0.015002166852355003, + 0.007010819390416145, + 0.019945133477449417, + 0.03423016890883446, + 0.006511680781841278, + 0.011760232970118523, + 0.009748483076691628, + 0.009803187102079391, + 0.04525408148765564, + 0.013865137472748756, + 0.06824091821908951, + 0.007758776191622019, + 0.0037329907063394785, + 0.006164638325572014, + 0.007038590963929892, + 0.005234848242253065, + 0.013373089954257011, + 0.008430096320807934, + 0.018311038613319397, + 0.020255960524082184, + 0.005172562785446644, + 0.014464007690548897, + 0.008469564840197563, + 0.008209277875721455, + 0.01938927173614502, + 0.010329811833798885, + 0.02555781416594982, + 0.022024307399988174, + 0.004195649642497301, + 0.010998581536114216, + 0.00713954446837306, + 0.008135431446135044, + 0.021930908784270287, + 0.007110216189175844, + 0.024867834523320198, + 0.008307899348437786, + 0.0018715898040682077, + 0.0053889332339167595, + 0.004628898575901985, + 0.005151752382516861, + 0.010907579213380814, + 0.005514106247574091, + 0.02118520252406597, + 0.008270765654742718, + 0.0019528599223122, + 0.0031482393387705088, + 0.0038433631416410208, + 0.005498449318110943, + 0.006572800222784281, + 0.003966501448303461, + 0.01039126981049776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.03372875973582268 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9103384017944336 + ], + "opposite_logprob_tokens": [ + -0.16033844649791718 + ], + "logprob_reward": -1.9103384017944336, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103384017944336, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1816956251859665, + 0.10222051292657852, + 0.11444493383169174, + 0.07354342192411423, + 0.12953676283359528, + 0.1547989398241043, + 0.1650882363319397, + 0.018343694508075714, + 0.02836182899773121, + 0.03196609392762184 + ], + "credit_map": [ + 0.013646180741488934, + 0.015358561649918556, + 0.01695452257990837, + 0.020767953246831894, + 0.01028632465749979, + 0.03618048503994942, + 0.020037245005369186, + 0.048464346677064896, + 0.0063514565117657185, + 0.018052933737635612, + 0.010994435288012028, + 0.010981936007738113, + 0.009374679066240788, + 0.015838541090488434, + 0.009898010641336441, + 0.02072850428521633, + 0.009163649752736092, + 0.013726773671805859, + 0.016964465379714966, + 0.01097146701067686, + 0.009985185228288174, + 0.018999449908733368, + 0.013014898635447025, + 0.0216190367937088, + 0.0071809785440564156, + 0.008264792151749134, + 0.008749594911932945, + 0.009399289265275002, + 0.0055032456293702126, + 0.011255016550421715, + 0.009830698370933533, + 0.01335980650037527, + 0.015563136897981167, + 0.007253644987940788, + 0.011610468849539757, + 0.009188521653413773, + 0.007678155321627855, + 0.029614659026265144, + 0.014201309531927109, + 0.03442685306072235, + 0.01869279146194458, + 0.0055117495357990265, + 0.009327949024736881, + 0.010455019772052765, + 0.007154403254389763, + 0.039804600179195404, + 0.015297388657927513, + 0.04855502396821976, + 0.02000320516526699, + 0.010902662761509418, + 0.025277074426412582, + 0.01644800417125225, + 0.017337318509817123, + 0.024313073605298996, + 0.01954358071088791, + 0.031263317912817, + 0.0034779186826199293, + 0.0007181851542554796, + 0.0018769908929243684, + 0.0014178541023284197, + 0.0019214224303141236, + 0.0028993638698011637, + 0.0015975842252373695, + 0.0044343736954033375, + 0.004418765660375357, + 0.0009448354248888791, + 0.0024185017682611942, + 0.001889198087155819, + 0.0028480703476816416, + 0.005168708506971598, + 0.001990295946598053, + 0.008683448657393456, + 0.004437234252691269, + 0.0013376943534240127, + 0.002736873459070921, + 0.00288599100895226, + 0.003611969295889139, + 0.005982960574328899, + 0.0025955187156796455, + 0.008377850987017155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -1.9103384017944336 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10035194456577301 + ], + "opposite_logprob_tokens": [ + -2.3503520488739014 + ], + "logprob_reward": -0.10035194456577301, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10035194456577301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13983482122421265, + 0.07908179610967636, + 0.11762985587120056, + 0.11693327128887177, + 0.06376379728317261, + 0.1191430315375328, + 0.19943907856941223, + 0.05101878568530083, + 0.04278523474931717, + 0.07037031650543213 + ], + "credit_map": [ + 0.011497072875499725, + 0.010908576659858227, + 0.016609840095043182, + 0.015277317725121975, + 0.006868750788271427, + 0.027619317173957825, + 0.015366069041192532, + 0.0356878861784935, + 0.0073027340695261955, + 0.009340781718492508, + 0.0070748114958405495, + 0.006820630747824907, + 0.006711319088935852, + 0.016004713252186775, + 0.007510598748922348, + 0.018316203728318214, + 0.013349564746022224, + 0.016983410343527794, + 0.011744911782443523, + 0.011036321520805359, + 0.009892292320728302, + 0.022073805332183838, + 0.011378309689462185, + 0.02117123268544674, + 0.0142919747158885, + 0.012521759606897831, + 0.017807183787226677, + 0.011785143055021763, + 0.008316688239574432, + 0.01999075524508953, + 0.01376801822334528, + 0.01845175214111805, + 0.010462380014359951, + 0.0062880199402570724, + 0.004463924095034599, + 0.005742254666984081, + 0.004952107090502977, + 0.010262753814458847, + 0.005939251743257046, + 0.015653109177947044, + 0.012906690128147602, + 0.007052783388644457, + 0.01466100849211216, + 0.010998960584402084, + 0.008220762014389038, + 0.021854765713214874, + 0.014530301094055176, + 0.02891775779426098, + 0.02620372548699379, + 0.01303131878376007, + 0.042620737105607986, + 0.01972818374633789, + 0.015160780400037766, + 0.02908669039607048, + 0.028188757598400116, + 0.025418894365429878, + 0.012260274961590767, + 0.0018009268678724766, + 0.005662752315402031, + 0.0037977665197104216, + 0.004339228849858046, + 0.008537858724594116, + 0.004428226035088301, + 0.010191753506660461, + 0.007565290667116642, + 0.0024775902274996042, + 0.005043619312345982, + 0.005254408344626427, + 0.005605608690530062, + 0.0044896225444972515, + 0.006263725459575653, + 0.006085368804633617, + 0.009914480149745941, + 0.0032362593337893486, + 0.005212758667767048, + 0.005122685804963112, + 0.008102410472929478, + 0.011011713184416294, + 0.0056205978617072105, + 0.022149410098791122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.10035194456577301 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234090685844421 + ], + "opposite_logprob_tokens": [ + -0.8984090685844421 + ], + "logprob_reward": -0.5234090685844421, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234090685844421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07715718448162079, + 0.03307744488120079, + 0.0583297498524189, + 0.05604118853807449, + 0.03948783501982689, + 0.3409753143787384, + 0.07610977441072464, + 0.07232772558927536, + 0.15990819036960602, + 0.08658545464277267 + ], + "credit_map": [ + 0.006316641345620155, + 0.005674288608133793, + 0.005871528759598732, + 0.007616291753947735, + 0.0033556073904037476, + 0.017643224447965622, + 0.0037873308174312115, + 0.0268922857940197, + 0.0021614895667880774, + 0.0024371773470193148, + 0.0033235105220228434, + 0.003174422075971961, + 0.0028749245684593916, + 0.006494705565273762, + 0.0038052010349929333, + 0.008806021884083748, + 0.004897498060017824, + 0.004795163869857788, + 0.008282193914055824, + 0.0053611453622579575, + 0.005131053738296032, + 0.010369696654379368, + 0.006492084823548794, + 0.013000928796827793, + 0.0060996063984930515, + 0.00422001862898469, + 0.0052375467494130135, + 0.004077457357198, + 0.003959983587265015, + 0.010887185111641884, + 0.005084117874503136, + 0.016475282609462738, + 0.005031811539083719, + 0.0033750031143426895, + 0.0047941687516868114, + 0.002895304700359702, + 0.003263317048549652, + 0.006173883099108934, + 0.004265810362994671, + 0.009688546881079674, + 0.057297270745038986, + 0.013955270871520042, + 0.025741832330822945, + 0.02601676620543003, + 0.025265347212553024, + 0.0732007697224617, + 0.04311862960457802, + 0.07637950032949448, + 0.010970472358167171, + 0.004597419872879982, + 0.006583661772310734, + 0.007065746001899242, + 0.007710673380643129, + 0.012663700617849827, + 0.005902323871850967, + 0.02061579003930092, + 0.016149066388607025, + 0.002644532360136509, + 0.006557404529303312, + 0.005081904586404562, + 0.007898101583123207, + 0.013294106349349022, + 0.004299596417695284, + 0.016403034329414368, + 0.022424863651394844, + 0.003062901319935918, + 0.009758777916431427, + 0.00823639240115881, + 0.018743079155683517, + 0.03244287520647049, + 0.00892265047878027, + 0.05631668120622635, + 0.013157776556909084, + 0.002482327166944742, + 0.006786659359931946, + 0.0059678032994270325, + 0.008415698073804379, + 0.014998502098023891, + 0.006772105116397142, + 0.028004605323076248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.5234090685844421 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in cinematic extreme close-up with a macro lens, the film features a slow and graceful 180-degree orbiting shot around a single candle flame. The candle, a simple column of white wax with a subtle dry-drip texture on its sides, rests on a dark, reflective wooden surface within a pitch-black room. Initially, its blackened wick supports only a feeble, sputtering flame—a tiny, almost translucent teardrop of pale yellow and blue light that flickers hesitantly, casting faint, trembling shadows. As the camera steadily orbits the subject, the flame undergoes a sudden, dramatic transformation; in a swift, decisive moment, it \"wakes up,\" instantly reaching upward and intensifying into a steady, confident, almond-shaped flame of brilliant orange-yellow, with a vibrant blue core at its base. This newly awakened flame now stands firm, radiating a powerful, warm, golden glow that pools on the candle's waxy surface and casts a soft, steady glow on the polished wood beneath, creating an atmosphere of deep intimacy and contemplation, completely focused on this singular moment of burning and its tranquil glow.", + "seed": 15264, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_14/index_6099_seed_15264_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0722462797537446, + "vlm_logprob": -0.0722462797537446, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.0722462797537446, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012279034592211246 + ], + "opposite_logprob_tokens": [ + -4.762279033660889 + ], + "logprob_reward": -0.012279034592211246, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012279034592211246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10223086178302765, + 0.09213000535964966, + 0.07516986131668091, + 0.17980730533599854, + 0.06696681678295135, + 0.12843391299247742, + 0.09627256542444229, + 0.09807609766721725, + 0.03738005831837654, + 0.1235324963927269 + ], + "credit_map": [ + 0.014557817950844765, + 0.013032342307269573, + 0.019533224403858185, + 0.023907095193862915, + 0.0074908616952598095, + 0.00787289533764124, + 0.00673662219196558, + 0.009100004099309444, + 0.012732432223856449, + 0.011831727810204029, + 0.015407238155603409, + 0.024546340107917786, + 0.005346407648175955, + 0.006777332630008459, + 0.007528705056756735, + 0.007959816604852676, + 0.01155890617519617, + 0.007251223549246788, + 0.012119712308049202, + 0.0150833111256361, + 0.005058355163782835, + 0.0076272268779575825, + 0.00821055006235838, + 0.0082605741918087, + 0.033074717968702316, + 0.01918984018266201, + 0.027885327115654945, + 0.02462179772555828, + 0.011404658667743206, + 0.018030688166618347, + 0.016141587868332863, + 0.029458703473210335, + 0.008428397588431835, + 0.005723710171878338, + 0.008956086821854115, + 0.011932815425097942, + 0.004344253335148096, + 0.007030152250081301, + 0.006290113087743521, + 0.014261282980442047, + 0.02128845825791359, + 0.014591776765882969, + 0.019140437245368958, + 0.01840074732899666, + 0.010674778372049332, + 0.01494098175317049, + 0.009368336759507656, + 0.020028406754136086, + 0.016395961865782738, + 0.01002766378223896, + 0.012327002361416817, + 0.016180813312530518, + 0.007788192946463823, + 0.011129346676170826, + 0.008103839121758938, + 0.014319736510515213, + 0.016887204721570015, + 0.010521036572754383, + 0.015382586978375912, + 0.015454688109457493, + 0.006408048793673515, + 0.011411171406507492, + 0.00931913685053587, + 0.012692224234342575, + 0.0060949623584747314, + 0.0032253412064164877, + 0.007074471563100815, + 0.00424250727519393, + 0.0029616050887852907, + 0.003931575920432806, + 0.003766673384234309, + 0.006082924082875252, + 0.01619279384613037, + 0.010140210390090942, + 0.023802725598216057, + 0.01417471282184124, + 0.009863318875432014, + 0.016137612983584404, + 0.01409279927611351, + 0.01912831701338291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 9 + ], + "selected_reward": -0.012279034592211246 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812277689576149 + ], + "opposite_logprob_tokens": [ + -3.2881228923797607 + ], + "logprob_reward": -0.03812277689576149, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812277689576149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20960880815982819, + 0.04574527591466904, + 0.0353151373565197, + 0.027114426717162132, + 0.08809450268745422, + 0.02461102604866028, + 0.08819636702537537, + 0.0492902435362339, + 0.3331235945224762, + 0.09890066832304001 + ], + "credit_map": [ + 0.027424294501543045, + 0.028215890750288963, + 0.030735768377780914, + 0.05398130416870117, + 0.01421044860035181, + 0.014360099099576473, + 0.022661274299025536, + 0.018019704148173332, + 0.005700657144188881, + 0.004116296302527189, + 0.006745830178260803, + 0.008476223796606064, + 0.004606869071722031, + 0.0040700421668589115, + 0.007229614071547985, + 0.004799745045602322, + 0.003945362754166126, + 0.003956938628107309, + 0.0050549195148050785, + 0.0067221433855593204, + 0.002321629086509347, + 0.0033273070584982634, + 0.004067969974130392, + 0.005918866489082575, + 0.004324170295149088, + 0.002783523639664054, + 0.004037119913846254, + 0.00422449829056859, + 0.0019329240312799811, + 0.0026552958879619837, + 0.0026848374400287867, + 0.004472058266401291, + 0.012091833166778088, + 0.007989149540662766, + 0.01306704431772232, + 0.011260775849223137, + 0.004569948185235262, + 0.009425375610589981, + 0.007869594730436802, + 0.021820785477757454, + 0.004342354368418455, + 0.002136436989530921, + 0.0035960394889116287, + 0.004228643141686916, + 0.002115418901666999, + 0.0019158209906890988, + 0.0022660810500383377, + 0.004010233096778393, + 0.017332563176751137, + 0.006510636303573847, + 0.015553900972008705, + 0.01461170706897974, + 0.00596938980743289, + 0.009696935303509235, + 0.006753132212907076, + 0.011768100783228874, + 0.006913539953529835, + 0.005526222288608551, + 0.006843243725597858, + 0.007534791249781847, + 0.0046210442669689655, + 0.004846092779189348, + 0.005261868704110384, + 0.007743440102785826, + 0.052271999418735504, + 0.02601272240281105, + 0.05751335993409157, + 0.0322830006480217, + 0.012216008268296719, + 0.04720962420105934, + 0.022892093285918236, + 0.08272479474544525, + 0.012620903551578522, + 0.008234571665525436, + 0.013906457461416721, + 0.013800657354295254, + 0.007748677860945463, + 0.01735921949148178, + 0.008571546524763107, + 0.016658633947372437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.03812277689576149 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281636476516724 + ], + "opposite_logprob_tokens": [ + -2.0178163051605225 + ], + "logprob_reward": -0.14281636476516724, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281636476516724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13985297083854675, + 0.04743324965238571, + 0.0399986170232296, + 0.046137019991874695, + 0.12635363638401031, + 0.0535019189119339, + 0.14084716141223907, + 0.32247045636177063, + 0.04258974269032478, + 0.04081523418426514 + ], + "credit_map": [ + 0.016040336340665817, + 0.01739424839615822, + 0.017281001433730125, + 0.03134864568710327, + 0.011898104101419449, + 0.015875693410634995, + 0.014079795219004154, + 0.015935156494379044, + 0.007443473674356937, + 0.003886668011546135, + 0.0077790808863937855, + 0.006670729257166386, + 0.003682230832055211, + 0.005379179958254099, + 0.005092991515994072, + 0.0074988966807723045, + 0.006036216858774424, + 0.004947846755385399, + 0.005264519713819027, + 0.005520577542483807, + 0.0029956293292343616, + 0.0037682910915464163, + 0.004324514418840408, + 0.00714101642370224, + 0.006213304121047258, + 0.004849981050938368, + 0.0076630981639027596, + 0.011110738851130009, + 0.00244370405562222, + 0.004358595237135887, + 0.0039429133757948875, + 0.005554689560085535, + 0.017172973603010178, + 0.010028217919170856, + 0.018414733931422234, + 0.018786702305078506, + 0.005628282204270363, + 0.01333850622177124, + 0.011240356601774693, + 0.0317438580095768, + 0.010602565482258797, + 0.005083034746348858, + 0.008408336900174618, + 0.008352554403245449, + 0.004067212343215942, + 0.004483319818973541, + 0.005507291294634342, + 0.0069976020604372025, + 0.024167589843273163, + 0.01242914143949747, + 0.023945637047290802, + 0.02371315099298954, + 0.009353496134281158, + 0.01676546409726143, + 0.015339878387749195, + 0.015132811851799488, + 0.05926813930273056, + 0.03302038833498955, + 0.04072580859065056, + 0.05278050899505615, + 0.023008478805422783, + 0.03271888568997383, + 0.026890968903899193, + 0.05405727028846741, + 0.006581366993486881, + 0.004497426096349955, + 0.01004689559340477, + 0.006589571945369244, + 0.0036043119616806507, + 0.00408522691577673, + 0.0033523349557071924, + 0.0038326107896864414, + 0.005133138038218021, + 0.004849172197282314, + 0.005190807394683361, + 0.0048355781473219395, + 0.003648703917860985, + 0.007557113654911518, + 0.004218915477395058, + 0.005381805822253227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.14281636476516724 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898040860891342 + ], + "opposite_logprob_tokens": [ + -2.5789804458618164 + ], + "logprob_reward": -0.07898040860891342, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898040860891342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10720822215080261, + 0.07122450321912766, + 0.15630634129047394, + 0.06218337267637253, + 0.1492890864610672, + 0.06159360706806183, + 0.1227446123957634, + 0.19044116139411926, + 0.04795089736580849, + 0.031058194115757942 + ], + "credit_map": [ + 0.017616698518395424, + 0.013094647787511349, + 0.016456296667456627, + 0.016680698841810226, + 0.008627025410532951, + 0.01157976035028696, + 0.009579990990459919, + 0.01357310451567173, + 0.009428306482732296, + 0.00783819705247879, + 0.010057703591883183, + 0.016745632514357567, + 0.00645456975325942, + 0.00602022185921669, + 0.00767134502530098, + 0.007008526939898729, + 0.018641779199242592, + 0.01722649857401848, + 0.021554354578256607, + 0.03202170133590698, + 0.008743282407522202, + 0.012339601293206215, + 0.018436208367347717, + 0.02734292484819889, + 0.009118590503931046, + 0.005360461305826902, + 0.00795658864080906, + 0.009994394145905972, + 0.0056595830246806145, + 0.0060663968324661255, + 0.006467848550528288, + 0.011559506878256798, + 0.021643051877617836, + 0.012172107584774494, + 0.020539892837405205, + 0.022080710157752037, + 0.010699638165533543, + 0.015692075714468956, + 0.011958274990320206, + 0.03450333699584007, + 0.010545400902628899, + 0.00690881535410881, + 0.009468222968280315, + 0.010052933357656002, + 0.005003972910344601, + 0.0043273321352899075, + 0.005894006695598364, + 0.009392921812832355, + 0.022777246311306953, + 0.011786372400820255, + 0.0194899532943964, + 0.017709428444504738, + 0.009315201081335545, + 0.017504557967185974, + 0.009781144559383392, + 0.014380707405507565, + 0.03616504743695259, + 0.02345881797373295, + 0.02627449482679367, + 0.02960575371980667, + 0.012565601617097855, + 0.016721108928322792, + 0.015391991473734379, + 0.03025834634900093, + 0.007413176819682121, + 0.005159626714885235, + 0.010528183542191982, + 0.00558991776779294, + 0.0028502121567726135, + 0.005071947816759348, + 0.004520682152360678, + 0.006817150395363569, + 0.004376881755888462, + 0.00340342428535223, + 0.004139116499572992, + 0.0027626596856862307, + 0.0019215535139665008, + 0.006625960115343332, + 0.003135740291327238, + 0.004692856688052416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.07898040860891342 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903281390666962 + ], + "opposite_logprob_tokens": [ + -2.4640328884124756 + ], + "logprob_reward": -0.08903281390666962, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903281390666962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1975545883178711, + 0.21814373135566711, + 0.15119758248329163, + 0.04209998622536659, + 0.059590600430965424, + 0.15232837200164795, + 0.05282846465706825, + 0.0208811666816473, + 0.030332263559103012, + 0.07504327595233917 + ], + "credit_map": [ + 0.025604059919714928, + 0.020507097244262695, + 0.02990233525633812, + 0.051968976855278015, + 0.014946079812943935, + 0.01728387549519539, + 0.02093859761953354, + 0.01640358194708824, + 0.03007553331553936, + 0.027458176016807556, + 0.03274386748671532, + 0.060632068663835526, + 0.01014281902462244, + 0.016196703538298607, + 0.019914573058485985, + 0.0209799911826849, + 0.026341719552874565, + 0.017992625012993813, + 0.025588996708393097, + 0.03534317761659622, + 0.008785170502960682, + 0.011147281154990196, + 0.011392656713724136, + 0.01460596639662981, + 0.005349195562303066, + 0.003385931486263871, + 0.00650967238470912, + 0.006377240177243948, + 0.0037959699984639883, + 0.004561341367661953, + 0.004994717426598072, + 0.007125916425138712, + 0.009338933043181896, + 0.0070223119109869, + 0.008829925209283829, + 0.009980758652091026, + 0.00396617129445076, + 0.005378853529691696, + 0.006263888441026211, + 0.008809755556285381, + 0.029012225568294525, + 0.016883138567209244, + 0.024919336661696434, + 0.020052414387464523, + 0.012218905612826347, + 0.013039221987128258, + 0.012354863807559013, + 0.023848265409469604, + 0.01000953372567892, + 0.004165363498032093, + 0.010834299959242344, + 0.008398185484111309, + 0.0030638272874057293, + 0.005718331318348646, + 0.0036266674287617207, + 0.007012249436229467, + 0.003146931529045105, + 0.0021199495531618595, + 0.002947029657661915, + 0.002806663978844881, + 0.002025526249781251, + 0.0026830974966287613, + 0.002371999667957425, + 0.0027799694798886776, + 0.004723803605884314, + 0.003315743524581194, + 0.004767449572682381, + 0.003317631082609296, + 0.0017674880800768733, + 0.004203140735626221, + 0.0021247707773, + 0.006112234201282263, + 0.010824359953403473, + 0.008374564349651337, + 0.011977453716099262, + 0.008626146242022514, + 0.004572957754135132, + 0.014758924022316933, + 0.005981245078146458, + 0.009927620179951191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.08903281390666962 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in cinematic extreme close-up with a macro lens, the film features a slow and graceful 180-degree orbiting shot around a single candle flame. The candle, a simple column of white wax with a subtle dry-drip texture on its sides, rests on a dark, reflective wooden surface within a pitch-black room. Initially, its blackened wick supports only a feeble, sputtering flame—a tiny, almost translucent teardrop of pale yellow and blue light that flickers hesitantly, casting faint, trembling shadows. As the camera steadily orbits the subject, the flame undergoes a sudden, dramatic transformation; in a swift, decisive moment, it \"wakes up,\" instantly reaching upward and intensifying into a steady, confident, almond-shaped flame of brilliant orange-yellow, with a vibrant blue core at its base. This newly awakened flame now stands firm, radiating a powerful, warm, golden glow that pools on the candle's waxy surface and casts a soft, steady glow on the polished wood beneath, creating an atmosphere of deep intimacy and contemplation, completely focused on this singular moment of burning and its tranquil glow.", + "seed": 15265, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_14/index_6099_seed_15265_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.8663836896419526, + "vlm_logprob": -0.8663836896419526, + "vlm_yes_no_margin": 1.175, + "vlm_token_credit": -0.8663836896419526, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07854542136192322 + ], + "opposite_logprob_tokens": [ + -2.703545331954956 + ], + "logprob_reward": -0.07854542136192322, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07854542136192322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11170163750648499, + 0.05555281788110733, + 0.07551951706409454, + 0.07467621564865112, + 0.27143731713294983, + 0.03892558813095093, + 0.18788129091262817, + 0.0980202853679657, + 0.07171501219272614, + 0.0145703861489892 + ], + "credit_map": [ + 0.010464765131473541, + 0.005328832659870386, + 0.0073075140826404095, + 0.005341002717614174, + 0.035319384187459946, + 0.03042745776474476, + 0.007180335000157356, + 0.010332341305911541, + 0.007983631454408169, + 0.005837851669639349, + 0.0033211070112884045, + 0.004494254477322102, + 0.011862969025969505, + 0.012461602687835693, + 0.004332993645220995, + 0.005258405581116676, + 0.011227455921471119, + 0.004753632936626673, + 0.0037651422899216413, + 0.0036157327704131603, + 0.009449263103306293, + 0.013131937012076378, + 0.005707158241420984, + 0.023869195953011513, + 0.008865537121891975, + 0.002968242857605219, + 0.006452924571931362, + 0.0038145084399729967, + 0.00631957221776247, + 0.012576502747833729, + 0.00729399174451828, + 0.026384931057691574, + 0.050062086433172226, + 0.009049709886312485, + 0.033138275146484375, + 0.01869688369333744, + 0.03544742241501808, + 0.04370209947228432, + 0.019961973652243614, + 0.06137885898351669, + 0.004467551596462727, + 0.0016722682630643249, + 0.003969961777329445, + 0.002457070630043745, + 0.004603350069373846, + 0.007253001444041729, + 0.002466699108481407, + 0.012035682797431946, + 0.042212169617414474, + 0.004986237268894911, + 0.01896575279533863, + 0.0053622000850737095, + 0.017041591927409172, + 0.06464721262454987, + 0.005942527670413256, + 0.028723591938614845, + 0.015172802843153477, + 0.004807517398148775, + 0.009553799405694008, + 0.006411574315279722, + 0.010576453059911728, + 0.024860475212335587, + 0.005488216411322355, + 0.02114945277571678, + 0.018073974177241325, + 0.003266291692852974, + 0.00724681606516242, + 0.00358539167791605, + 0.006379154045134783, + 0.01804034598171711, + 0.003907614853233099, + 0.01121542602777481, + 0.002454079920426011, + 0.000898639322258532, + 0.001780951046384871, + 0.0012481671292334795, + 0.0015943977050483227, + 0.0023895702324807644, + 0.0013862716732546687, + 0.0028183094691485167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.07854542136192322 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029822736978530884 + ], + "opposite_logprob_tokens": [ + -3.529822826385498 + ], + "logprob_reward": -0.029822736978530884, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029822736978530884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07051453739404678, + 0.06336871534585953, + 0.07338070869445801, + 0.0556219257414341, + 0.4766511917114258, + 0.09804271161556244, + 0.04805394262075424, + 0.030338430777192116, + 0.011581886559724808, + 0.07244588434696198 + ], + "credit_map": [ + 0.008255253545939922, + 0.008082869462668896, + 0.00619654031470418, + 0.006134901661425829, + 0.012604386545717716, + 0.014222254045307636, + 0.006789522711187601, + 0.008228814229369164, + 0.008512918837368488, + 0.004724087193608284, + 0.0030221419874578714, + 0.0029151637572795153, + 0.018067510798573494, + 0.018298910930752754, + 0.003038071794435382, + 0.00478991586714983, + 0.011338859796524048, + 0.002827842254191637, + 0.004297125618904829, + 0.002898815553635359, + 0.008881943300366402, + 0.013874508440494537, + 0.003798832418397069, + 0.025462789461016655, + 0.007047008723020554, + 0.002402820624411106, + 0.005583146587014198, + 0.003628432983532548, + 0.005256862845271826, + 0.008541152812540531, + 0.005903121083974838, + 0.01725938729941845, + 0.10617527365684509, + 0.017778897657990456, + 0.03587580844759941, + 0.01824801042675972, + 0.06520415097475052, + 0.0833907425403595, + 0.0313761830329895, + 0.11860214173793793, + 0.009959055110812187, + 0.003015464171767235, + 0.011325069703161716, + 0.004678295459598303, + 0.011863150633871555, + 0.01831694133579731, + 0.008449004963040352, + 0.030435746535658836, + 0.007645559962838888, + 0.002411080291494727, + 0.010764081962406635, + 0.002536109881475568, + 0.004764581564813852, + 0.005430841818451881, + 0.0028269090689718723, + 0.011674783192574978, + 0.0044845351949334145, + 0.0015485654585063457, + 0.0055594677105546, + 0.0016232725465670228, + 0.003060418413951993, + 0.006101179402321577, + 0.0020663421601057053, + 0.00589465256780386, + 0.0017726519145071507, + 0.0010832283878698945, + 0.0013256327947601676, + 0.0007580430246889591, + 0.001115669496357441, + 0.0024178004823625088, + 0.0006443269667215645, + 0.0024645342491567135, + 0.012624301016330719, + 0.003256307914853096, + 0.007694210857152939, + 0.0040467665530741215, + 0.014928239397704601, + 0.010439704172313213, + 0.006258274894207716, + 0.01319807954132557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.029822736978530884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -3.65132999420166 + ], + "opposite_logprob_tokens": [ + -0.02632994018495083 + ], + "logprob_reward": -3.65132999420166, + "yes_no_margin_reward": -3.625, + "token_credit_reward": -3.65132999420166, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.6905257105827332, + 0.03416101634502411, + 0.04286346212029457, + 0.04410263150930405, + 0.0585678406059742, + 0.03421735763549805, + 0.04686163738369942, + 0.027549991384148598, + 0.007397996261715889, + 0.013752332888543606 + ], + "credit_map": [ + 0.009429382160305977, + 0.00761025957763195, + 0.0033393269404768944, + 0.005850175395607948, + 0.32352790236473083, + 0.3155268430709839, + 0.006217723712325096, + 0.01902407966554165, + 0.002756516681984067, + 0.002045764122158289, + 0.001246932428330183, + 0.001486082561314106, + 0.010928097181022167, + 0.011719858273863792, + 0.0019703225698322058, + 0.002007444854825735, + 0.006227385252714157, + 0.0025637303479015827, + 0.003037749556824565, + 0.0022913850843906403, + 0.00534806028008461, + 0.0070966556668281555, + 0.003112402278929949, + 0.013186096213757992, + 0.007546903099864721, + 0.001796485623344779, + 0.0030794448684901, + 0.001768201356753707, + 0.004328400827944279, + 0.007711953949183226, + 0.0033666114322841167, + 0.01450463104993105, + 0.014446273446083069, + 0.002295341109856963, + 0.005159582942724228, + 0.0024810025934129953, + 0.011236688122153282, + 0.009484452195465565, + 0.0027857422828674316, + 0.010678757913410664, + 0.00498645706102252, + 0.0022411770187318325, + 0.0037336384411901236, + 0.0020755308214575052, + 0.004560243804007769, + 0.005790829658508301, + 0.0025733427610248327, + 0.008256137371063232, + 0.00775685952976346, + 0.0018745868001133204, + 0.00298236683011055, + 0.0013653726782649755, + 0.00525378342717886, + 0.01897072046995163, + 0.0016412509139627218, + 0.00701669417321682, + 0.003926515579223633, + 0.0015323812840506434, + 0.004983040038496256, + 0.0019612221512943506, + 0.0033718508202582598, + 0.006198473274707794, + 0.0018596663139760494, + 0.0037168434355407953, + 0.0011586989276111126, + 0.0002841118548531085, + 0.0010682063875719905, + 0.0005929159233346581, + 0.0006313027697615325, + 0.0012100385501980782, + 0.0005317030590958893, + 0.0019210187019780278, + 0.0015901856822893023, + 0.0008823712123557925, + 0.0013685023877769709, + 0.0009036397677846253, + 0.0015581031329929829, + 0.0022601457312703133, + 0.001032053492963314, + 0.0041573322378098965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -3.65132999420166 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048725515604019165 + ], + "opposite_logprob_tokens": [ + -3.0487256050109863 + ], + "logprob_reward": -0.048725515604019165, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048725515604019165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3033490777015686, + 0.07444030791521072, + 0.059849996119737625, + 0.12356030195951462, + 0.09996441751718521, + 0.06968246400356293, + 0.1643158197402954, + 0.0629214495420456, + 0.013442492112517357, + 0.02847364917397499 + ], + "credit_map": [ + 0.007807672489434481, + 0.004986323416233063, + 0.005124800372868776, + 0.0036818727385252714, + 0.13394786417484283, + 0.13218308985233307, + 0.006219150964170694, + 0.009398303925991058, + 0.006949357222765684, + 0.005151850171387196, + 0.0034142823424190283, + 0.004867332521826029, + 0.02132193185389042, + 0.02166314609348774, + 0.0054555912502110004, + 0.005616815760731697, + 0.008512746542692184, + 0.006364953238517046, + 0.004950404167175293, + 0.004932781681418419, + 0.007604324258863926, + 0.008827460929751396, + 0.005964546464383602, + 0.012692778371274471, + 0.021667741239070892, + 0.004833833314478397, + 0.010261701419949532, + 0.005506595596671104, + 0.011287468485534191, + 0.021717041730880737, + 0.008568964898586273, + 0.03971695899963379, + 0.022845007479190826, + 0.003660141723230481, + 0.0078111193142831326, + 0.003861465724185109, + 0.022886143997311592, + 0.015712155029177666, + 0.004595629405230284, + 0.018592752516269684, + 0.009839453734457493, + 0.0035443813540041447, + 0.007409386802464724, + 0.003729018848389387, + 0.009137650951743126, + 0.012507541105151176, + 0.00515333516523242, + 0.01836169697344303, + 0.03555983677506447, + 0.005272220820188522, + 0.005988811608403921, + 0.003128946991637349, + 0.01811155118048191, + 0.07097188383340836, + 0.0032951361499726772, + 0.02198743261396885, + 0.009563281200826168, + 0.00282606715336442, + 0.01169566996395588, + 0.003402763744816184, + 0.006167079322040081, + 0.014698775485157967, + 0.003789832117035985, + 0.010777981020510197, + 0.0027958666905760765, + 0.0008371557341888547, + 0.0015100938035175204, + 0.0010339899454265833, + 0.001127558178268373, + 0.0027912866789847612, + 0.0008285724325105548, + 0.0025179684162139893, + 0.005011455621570349, + 0.0021215167362242937, + 0.002375943586230278, + 0.0018713439349085093, + 0.004587869625538588, + 0.004036340396851301, + 0.0021590320393443108, + 0.006310145370662212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.048725515604019165 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234947800636292 + ], + "opposite_logprob_tokens": [ + -0.8984947800636292 + ], + "logprob_reward": -0.5234947800636292, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234947800636292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4104993939399719, + 0.1608601212501526, + 0.09068711847066879, + 0.022179586812853813, + 0.08350568264722824, + 0.05234220623970032, + 0.06045398488640785, + 0.10156550258398056, + 0.005706892814487219, + 0.012199502438306808 + ], + "credit_map": [ + 0.006901396904140711, + 0.004849668592214584, + 0.004324792418628931, + 0.00417292257770896, + 0.1885356456041336, + 0.18533501029014587, + 0.004685144405812025, + 0.011694824323058128, + 0.012387379072606564, + 0.006206733640283346, + 0.004526765551418066, + 0.004444479010999203, + 0.05830247327685356, + 0.06136176362633705, + 0.003638055408373475, + 0.009992473758757114, + 0.013429934158921242, + 0.0036459572147578, + 0.004479642957448959, + 0.002854304388165474, + 0.011721083894371986, + 0.017164865508675575, + 0.005076576489955187, + 0.03231475502252579, + 0.004077669233083725, + 0.001150256721302867, + 0.001664106035605073, + 0.0007990084122866392, + 0.002296897117048502, + 0.0037156599573791027, + 0.0016477120807394385, + 0.006828275974839926, + 0.020010193809866905, + 0.0032440887298434973, + 0.006375537719577551, + 0.0031692509073764086, + 0.013131240382790565, + 0.014853241853415966, + 0.003926459699869156, + 0.0187956765294075, + 0.007329301908612251, + 0.0026035085320472717, + 0.004445740487426519, + 0.0026894165202975273, + 0.011324407532811165, + 0.009319785051047802, + 0.0034049490932375193, + 0.011225101538002491, + 0.011377728544175625, + 0.0020574925001710653, + 0.005380640737712383, + 0.0015409382758662105, + 0.006397366523742676, + 0.023844590410590172, + 0.0016051934799179435, + 0.008250029757618904, + 0.016790298745036125, + 0.0036477502435445786, + 0.017272787168622017, + 0.00479397177696228, + 0.010405870154500008, + 0.02525797113776207, + 0.005324276629835367, + 0.018072573468089104, + 0.0012813113862648606, + 0.0003406764881219715, + 0.0006983547937124968, + 0.0005198264261707664, + 0.0004133880138397217, + 0.0009932712418958545, + 0.00046224077232182026, + 0.0009978236630558968, + 0.002346868859604001, + 0.0007830762770026922, + 0.001205109991133213, + 0.0007604199927300215, + 0.0024759217631071806, + 0.0017413182649761438, + 0.0010730782523751259, + 0.0018137100851163268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.5234947800636292 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in cinematic extreme close-up with a macro lens, the film features a slow and graceful 180-degree orbiting shot around a single candle flame. The candle, a simple column of white wax with a subtle dry-drip texture on its sides, rests on a dark, reflective wooden surface within a pitch-black room. Initially, its blackened wick supports only a feeble, sputtering flame—a tiny, almost translucent teardrop of pale yellow and blue light that flickers hesitantly, casting faint, trembling shadows. As the camera steadily orbits the subject, the flame undergoes a sudden, dramatic transformation; in a swift, decisive moment, it \"wakes up,\" instantly reaching upward and intensifying into a steady, confident, almond-shaped flame of brilliant orange-yellow, with a vibrant blue core at its base. This newly awakened flame now stands firm, radiating a powerful, warm, golden glow that pools on the candle's waxy surface and casts a soft, steady glow on the polished wood beneath, creating an atmosphere of deep intimacy and contemplation, completely focused on this singular moment of burning and its tranquil glow.", + "seed": 15266, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_14/index_6099_seed_15266_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.5224933810532093, + "vlm_logprob": -0.5224933810532093, + "vlm_yes_no_margin": 1.500000011920929, + "vlm_token_credit": -0.5224933810532093, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.44045746326446533 + ], + "opposite_logprob_tokens": [ + -1.0654574632644653 + ], + "logprob_reward": -0.44045746326446533, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.44045746326446533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3771427273750305, + 0.0425717867910862, + 0.1282716989517212, + 0.15723900496959686, + 0.0324518159031868, + 0.03855874389410019, + 0.035681918263435364, + 0.07764306664466858, + 0.07952801883220673, + 0.03091116063296795 + ], + "credit_map": [ + 0.01376201119273901, + 0.15342487394809723, + 0.020568175241351128, + 0.008790586143732071, + 0.01750924624502659, + 0.14308404922485352, + 0.007922003977000713, + 0.012081851251423359, + 0.005395081359893084, + 0.005210204515606165, + 0.00687449099496007, + 0.003412865335121751, + 0.006873561535030603, + 0.005059284158051014, + 0.0038535860367119312, + 0.0058927154168486595, + 0.024983618408441544, + 0.011052717454731464, + 0.031598303467035294, + 0.007774912286549807, + 0.0075803776271641254, + 0.018370699137449265, + 0.00752140861004591, + 0.019389662891626358, + 0.0198991559445858, + 0.017638860270380974, + 0.03511246666312218, + 0.016379863023757935, + 0.01023195218294859, + 0.026230433955788612, + 0.011611792258918285, + 0.020134489983320236, + 0.00384820019826293, + 0.002490786137059331, + 0.007707837503403425, + 0.003461257554590702, + 0.0016631463076919317, + 0.004244162235409021, + 0.003167372662574053, + 0.005869052838534117, + 0.004825504031032324, + 0.002083371626213193, + 0.012530602514743805, + 0.0036677629686892033, + 0.0017242489848285913, + 0.005803481675684452, + 0.003179705934599042, + 0.004744069650769234, + 0.005549852270632982, + 0.0018769040470942855, + 0.007897568866610527, + 0.0021514464169740677, + 0.0017556246602907777, + 0.005550248548388481, + 0.0028832347597926855, + 0.008017039857804775, + 0.012590555474162102, + 0.00410360936075449, + 0.015893487259745598, + 0.003167137037962675, + 0.005335073918104172, + 0.012147889472544193, + 0.004111101850867271, + 0.020294221118092537, + 0.014685857109725475, + 0.0023396292235702276, + 0.013775310479104519, + 0.0032559134997427464, + 0.003983693663030863, + 0.01303865946829319, + 0.004253220744431019, + 0.024195734411478043, + 0.003761894768103957, + 0.0009272897150367498, + 0.004703199490904808, + 0.0014878545189276338, + 0.0017339972546324134, + 0.005257685203105211, + 0.0016362466849386692, + 0.011402992531657219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.44045746326446533 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018210364505648613 + ], + "opposite_logprob_tokens": [ + -4.018210411071777 + ], + "logprob_reward": -0.018210364505648613, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018210364505648613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.536462128162384, + 0.046600863337516785, + 0.041199371218681335, + 0.05686037987470627, + 0.021643642336130142, + 0.011112063191831112, + 0.022513337433338165, + 0.04005945846438408, + 0.14343129098415375, + 0.08011749386787415 + ], + "credit_map": [ + 0.019958628341555595, + 0.21978048980236053, + 0.0359899178147316, + 0.010680950246751308, + 0.02653982862830162, + 0.20005089044570923, + 0.009808136150240898, + 0.01365326065570116, + 0.004685594700276852, + 0.004374941345304251, + 0.012960298918187618, + 0.003243589773774147, + 0.0034560475032776594, + 0.008284639567136765, + 0.004308394156396389, + 0.0052873557433485985, + 0.007680164650082588, + 0.0035613656509667635, + 0.007746008224785328, + 0.002795150503516197, + 0.0024234324228018522, + 0.007243460044264793, + 0.0028495739679783583, + 0.006900216452777386, + 0.006826197262853384, + 0.005958197638392448, + 0.01330366637557745, + 0.0063815345056355, + 0.003688566852360964, + 0.009275561198592186, + 0.004250720608979464, + 0.007175937760621309, + 0.0031398481223732233, + 0.0014907907461747527, + 0.005916288588196039, + 0.0023279946763068438, + 0.0011439460795372725, + 0.0032159853726625443, + 0.001989658223465085, + 0.0024191304109990597, + 0.001585570746101439, + 0.0005882370169274509, + 0.0034632578026503325, + 0.001081550377421081, + 0.000520217465236783, + 0.001757137244567275, + 0.0009252921445295215, + 0.0011908002197742462, + 0.003731534583494067, + 0.001052757608704269, + 0.005319252610206604, + 0.0014100434491410851, + 0.0009772730991244316, + 0.0037306034937500954, + 0.0014161583967506886, + 0.004875714424997568, + 0.005947110243141651, + 0.0016632506158202887, + 0.007204731926321983, + 0.002548885066062212, + 0.0024101664312183857, + 0.0064988029189407825, + 0.0025410784874111414, + 0.011245431378483772, + 0.020850135013461113, + 0.0035968623124063015, + 0.02653132751584053, + 0.006033072713762522, + 0.00752637954428792, + 0.023694951087236404, + 0.006933166645467281, + 0.04826539754867554, + 0.009685888886451721, + 0.0021336134523153305, + 0.012209235690534115, + 0.0037752490025013685, + 0.004400122445076704, + 0.014257156290113926, + 0.00542664946988225, + 0.02822958119213581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.018210364505648613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7015305757522583 + ], + "opposite_logprob_tokens": [ + -0.2015305757522583 + ], + "logprob_reward": -1.7015305757522583, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7015305757522583, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.49185848236083984, + 0.03745117038488388, + 0.0367901511490345, + 0.1107856035232544, + 0.018125809729099274, + 0.013390897773206234, + 0.025405248627066612, + 0.015234633348882198, + 0.2319418489933014, + 0.01901613548398018 + ], + "credit_map": [ + 0.01936386525630951, + 0.19422738254070282, + 0.03600964695215225, + 0.009941678494215012, + 0.022756479680538177, + 0.1816231906414032, + 0.008663407526910305, + 0.019272813573479652, + 0.0034510865807533264, + 0.004898982122540474, + 0.005399246700108051, + 0.0036196212749928236, + 0.004058604594320059, + 0.006786681246012449, + 0.0043725548312067986, + 0.004864389076828957, + 0.004083412699401379, + 0.0029928695876151323, + 0.009924089536070824, + 0.0029687839560210705, + 0.002685768296942115, + 0.0065455809235572815, + 0.0025753220543265343, + 0.005014324095100164, + 0.011511878110468388, + 0.010272148065268993, + 0.027742093428969383, + 0.011769559234380722, + 0.0070215933956205845, + 0.01920260116457939, + 0.008306714706122875, + 0.014959008432924747, + 0.002194951055571437, + 0.0012644294183701277, + 0.005118818953633308, + 0.001868441584520042, + 0.0010865413350984454, + 0.002532858867198229, + 0.0015872518997639418, + 0.0024725173134356737, + 0.0014665921917185187, + 0.0008794812019914389, + 0.0039005009457468987, + 0.0012730376329272985, + 0.0007855153526179492, + 0.0019959344062954187, + 0.001437522703781724, + 0.0016523138619959354, + 0.0037967103999108076, + 0.0014535757945850492, + 0.005448103416711092, + 0.0022798911668360233, + 0.0012571871047839522, + 0.004216141067445278, + 0.002105217892676592, + 0.004848421551287174, + 0.0017675438430160284, + 0.0007138049695640802, + 0.0039017696399241686, + 0.0009917889256030321, + 0.0010580226080492139, + 0.002512926934286952, + 0.001162869855761528, + 0.0031259062234312296, + 0.04420403763651848, + 0.006754529662430286, + 0.038067057728767395, + 0.009111057966947556, + 0.012026304379105568, + 0.03861159458756447, + 0.011338132433593273, + 0.07182914018630981, + 0.0022703497670590878, + 0.0010041373316198587, + 0.003929184749722481, + 0.0013046945678070188, + 0.0017458178335800767, + 0.0028380239382386208, + 0.0017513135680928826, + 0.004172613378614187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -1.7015305757522583 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023307522758841515 + ], + "opposite_logprob_tokens": [ + -3.7733075618743896 + ], + "logprob_reward": -0.023307522758841515, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023307522758841515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45511695742607117, + 0.09954244643449783, + 0.054388877004384995, + 0.04320789873600006, + 0.022397207096219063, + 0.03075767494738102, + 0.053499653935432434, + 0.042307298630476, + 0.16620247066020966, + 0.03257958963513374 + ], + "credit_map": [ + 0.017616629600524902, + 0.18075041472911835, + 0.029859228059649467, + 0.010945873335003853, + 0.022028256207704544, + 0.16657249629497528, + 0.010059958323836327, + 0.017284046858549118, + 0.009747829288244247, + 0.010576575063169003, + 0.01906740292906761, + 0.01267517451196909, + 0.008154005743563175, + 0.013192187063395977, + 0.014417051337659359, + 0.011712207458913326, + 0.008296169340610504, + 0.004574453458189964, + 0.016380922868847847, + 0.003426792798563838, + 0.0044374787248671055, + 0.007413698360323906, + 0.003635238390415907, + 0.006224119104444981, + 0.004437391646206379, + 0.0035896645858883858, + 0.010002522729337215, + 0.005621439311653376, + 0.0030781205277889967, + 0.0066936565563082695, + 0.0034952545538544655, + 0.006289845332503319, + 0.0024041261058300734, + 0.0012955579441040754, + 0.007518643978983164, + 0.002009003423154354, + 0.0010924020316451788, + 0.003584390040487051, + 0.001790071139112115, + 0.0027030122000724077, + 0.002975547220557928, + 0.001750518917106092, + 0.011162539944052696, + 0.003156327875331044, + 0.0012681405059993267, + 0.004799270536750555, + 0.0025058488827198744, + 0.0031394795514643192, + 0.008206251077353954, + 0.0024619163013994694, + 0.010698636062443256, + 0.003709565382450819, + 0.0020664602052420378, + 0.009538738057017326, + 0.003766339272260666, + 0.013051738031208515, + 0.006470819935202599, + 0.0024619109462946653, + 0.00834352895617485, + 0.00238122814334929, + 0.002604673383757472, + 0.006856759078800678, + 0.0026988007593899965, + 0.010489574633538723, + 0.015800556167960167, + 0.0076595027931034565, + 0.031435299664735794, + 0.01196694653481245, + 0.011269534938037395, + 0.026641400530934334, + 0.011113280430436134, + 0.050315938889980316, + 0.0034579134080559015, + 0.0014688113005831838, + 0.007808101363480091, + 0.00284812250174582, + 0.0025482086930423975, + 0.004799844231456518, + 0.0026957134250551462, + 0.006952871568500996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.023307522758841515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the candle a white wax column with a dry-drip texture resting on a dark reflective wooden surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a slow 180-degree orbiting shot around the candle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flame transform from a feeble sputtering pale yellow/blue flame to a steady brilliant orange-yellow flame with a blue core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a pitch-black room?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42896097898483276 + ], + "opposite_logprob_tokens": [ + -1.0539610385894775 + ], + "logprob_reward": -0.42896097898483276, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.42896097898483276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3366281986236572, + 0.05276178941130638, + 0.08055263012647629, + 0.030404355376958847, + 0.12328610569238663, + 0.18348358571529388, + 0.05259716138243675, + 0.0870179757475853, + 0.03663170337677002, + 0.016636541113257408 + ], + "credit_map": [ + 0.019662272185087204, + 0.10642437636852264, + 0.06205335631966591, + 0.01270291768014431, + 0.014837675727903843, + 0.09461084008216858, + 0.0085400165989995, + 0.01779673993587494, + 0.006089176516979933, + 0.005780471954494715, + 0.013618380762636662, + 0.003422743873670697, + 0.003639496862888336, + 0.0128659438341856, + 0.00301494006998837, + 0.004330637399107218, + 0.008077913895249367, + 0.006599990651011467, + 0.022647690027952194, + 0.008866023272275925, + 0.005442693363875151, + 0.012716598808765411, + 0.007067836821079254, + 0.009133880957961082, + 0.003377649001777172, + 0.0026868279092013836, + 0.007377564907073975, + 0.0044772126711905, + 0.0024804570712149143, + 0.0035882294178009033, + 0.00242203613743186, + 0.003994377329945564, + 0.013737799599766731, + 0.007551979273557663, + 0.03848736360669136, + 0.01490077469497919, + 0.006456308998167515, + 0.016846347600221634, + 0.009519161656498909, + 0.015786362811923027, + 0.024796640500426292, + 0.011119019240140915, + 0.05082992836833, + 0.018399013206362724, + 0.008354275487363338, + 0.025455936789512634, + 0.0149764409288764, + 0.02955232746899128, + 0.009247292764484882, + 0.003975220024585724, + 0.011994289234280586, + 0.003685412462800741, + 0.0028644984122365713, + 0.009825404733419418, + 0.00286543439142406, + 0.008139610290527344, + 0.013225063681602478, + 0.004427362699061632, + 0.017368068918585777, + 0.005195386242121458, + 0.005433484446257353, + 0.014195741154253483, + 0.0059371222741901875, + 0.021235737949609756, + 0.006527986843138933, + 0.001241636578924954, + 0.005458742380142212, + 0.0017382974037900567, + 0.002168103586882353, + 0.005971742793917656, + 0.002636328572407365, + 0.010888866148889065, + 0.0017919238889589906, + 0.0006142295314930379, + 0.0037801568396389484, + 0.0011438766960054636, + 0.0010986831039190292, + 0.002452561166137457, + 0.0012270796578377485, + 0.004528030287474394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.42896097898483276 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_15/index_tensor([6828, 6828])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_15/index_tensor([6828, 6828])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b57e158f6b12ba05d1bec8e7fd096832d2940f6c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_15/index_tensor([6828, 6828])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, high-angle drone shot unfolds beneath a cloudless, azure sky, capturing a seemingly boundless expanse of desert. The camera then executes a swift and graceful swoop, rapidly descending to the ground, its movement fluid and purposeful. As it skims mere inches above the surface, it carefully reveals intricate, wind-sculpted ripples in the sand, their delicate patterns catching the brilliant light. The camera glides past the occasional, sparse stand of saguaro cacti, their weathered, dark green, ribbed forms casting long, stark shadows across the golden ground. Without pausing, the camera steadily and majestically ascends, pulling back to a breathtaking aerial view that re-emphasizes the sheer, overwhelming scale of the arid landscape. The entire scene is bathed in the warm, brilliant glow of the afternoon sun, igniting the rolling dunes in rich hues of deep orange and rich ochre, stretching endlessly to a shimmering, heat-hazed horizon, creating a powerful atmosphere of tranquility, majesty, and beautiful desolation.", + "seed": 26236, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_15/index_6828_seed_26236_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.051506445021368565, + "vlm_logprob": -0.051506445021368565, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.051506445021368565, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015254301251843572 + ], + "opposite_logprob_tokens": [ + -7.251525402069092 + ], + "logprob_reward": -0.0015254301251843572, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0015254301251843572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3933553993701935, + 0.1077190712094307, + 0.18733848631381989, + 0.03866070136427879, + 0.024830877780914307, + 0.029662586748600006, + 0.0249489638954401, + 0.060734450817108154, + 0.04091273993253708, + 0.09183672070503235 + ], + "credit_map": [ + 0.02397107519209385, + 0.011851451359689236, + 0.059356220066547394, + 0.04230174422264099, + 0.018932996317744255, + 0.01585621014237404, + 0.12750183045864105, + 0.09358391165733337, + 0.007550744339823723, + 0.003916636109352112, + 0.029480455443263054, + 0.015296061523258686, + 0.0037663779221475124, + 0.0041742511093616486, + 0.02882291190326214, + 0.014711642637848854, + 0.014379165135324001, + 0.019280292093753815, + 0.022085092961788177, + 0.01689111813902855, + 0.009755631908774376, + 0.04334045946598053, + 0.016518859192728996, + 0.04508790373802185, + 0.0018848625477403402, + 0.0016156743513420224, + 0.004139917902648449, + 0.003960573114454746, + 0.0019541785586625338, + 0.0018224529922008514, + 0.011527291499078274, + 0.011755754239857197, + 0.002039731713011861, + 0.0026529289316385984, + 0.0018883716547861695, + 0.0028134447056800127, + 0.0016245375154539943, + 0.0022235331125557423, + 0.004896875936537981, + 0.006691457238048315, + 0.0020615009125322104, + 0.002443291014060378, + 0.003401614725589752, + 0.00448773754760623, + 0.002957349643111229, + 0.0037045644130557775, + 0.005965694785118103, + 0.004640836734324694, + 0.0022468254901468754, + 0.004950988572090864, + 0.003646503435447812, + 0.0032349370885640383, + 0.0020566226448863745, + 0.0026596507523208857, + 0.003081975970417261, + 0.003071462269872427, + 0.004757021088153124, + 0.009304446168243885, + 0.005523913074284792, + 0.009037752635776997, + 0.004265883471816778, + 0.006194941233843565, + 0.012963124550879002, + 0.008687368594110012, + 0.002844830509275198, + 0.004158122465014458, + 0.0027541606687009335, + 0.011089270003139973, + 0.0023481997195631266, + 0.004018201492726803, + 0.008986016735434532, + 0.004713939968496561, + 0.012477828189730644, + 0.009356009773910046, + 0.008618609979748726, + 0.009521868079900742, + 0.006973403971642256, + 0.02623691037297249, + 0.00823860615491867, + 0.010413493029773235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0015254301251843572 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784713387489319 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784713387489319, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784713387489319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3369385898113251, + 0.08163565397262573, + 0.12961901724338531, + 0.04041854292154312, + 0.055205170065164566, + 0.12293097376823425, + 0.0338970385491848, + 0.05540817975997925, + 0.09353622794151306, + 0.05041057989001274 + ], + "credit_map": [ + 0.008157634176313877, + 0.004119460936635733, + 0.025120694190263748, + 0.03139275684952736, + 0.008534085005521774, + 0.0070360926911234856, + 0.14157958328723907, + 0.11099830269813538, + 0.005241207778453827, + 0.0038310042582452297, + 0.022597728297114372, + 0.011335776187479496, + 0.004389030858874321, + 0.007388316094875336, + 0.01009017787873745, + 0.016762420535087585, + 0.0067853741347789764, + 0.00792321190237999, + 0.01412845030426979, + 0.012550934217870235, + 0.006906731054186821, + 0.024712178856134415, + 0.020110299810767174, + 0.03650182858109474, + 0.0031330042984336615, + 0.0024949628859758377, + 0.009014786221086979, + 0.009484429843723774, + 0.0026866872794926167, + 0.0026768879033625126, + 0.005958753637969494, + 0.004969033878296614, + 0.003095947904512286, + 0.0046903095208108425, + 0.006970015820115805, + 0.01146025862544775, + 0.0029885105323046446, + 0.005424280650913715, + 0.010011312551796436, + 0.010564533993601799, + 0.006729513872414827, + 0.01614055596292019, + 0.019386574625968933, + 0.023204950615763664, + 0.008707050234079361, + 0.015033984556794167, + 0.01682146079838276, + 0.016906889155507088, + 0.004159050993621349, + 0.009600108489394188, + 0.0029966498259454966, + 0.003813926363363862, + 0.0033386575523763895, + 0.0024692725855857134, + 0.00373212737031281, + 0.0037872479297220707, + 0.0034938696771860123, + 0.004432876128703356, + 0.005941616836935282, + 0.007608849089592695, + 0.003693137550726533, + 0.006046663969755173, + 0.010141165927052498, + 0.014049995690584183, + 0.0039684101939201355, + 0.006607573479413986, + 0.008398949168622494, + 0.020730186253786087, + 0.005720644257962704, + 0.0076298462226986885, + 0.026942552998661995, + 0.013538067229092121, + 0.005148708820343018, + 0.00645324494689703, + 0.005145894829183817, + 0.005861668381839991, + 0.005014124792069197, + 0.011863133870065212, + 0.006188899744302034, + 0.00473490497097373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.009784713387489319 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633641801774502 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633641801774502, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633641801774502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22620965540409088, + 0.09659384191036224, + 0.11770734190940857, + 0.13457730412483215, + 0.07364214956760406, + 0.10953640192747116, + 0.02762347087264061, + 0.07437532395124435, + 0.06953595578670502, + 0.07019858807325363 + ], + "credit_map": [ + 0.023043323308229446, + 0.015263316221535206, + 0.050781361758708954, + 0.03575865924358368, + 0.023080814629793167, + 0.022290170192718506, + 0.03045104444026947, + 0.025540977716445923, + 0.0042219688184559345, + 0.0044123344123363495, + 0.014252206310629845, + 0.021339019760489464, + 0.006725797429680824, + 0.006610760930925608, + 0.018930630758404732, + 0.02010112628340721, + 0.007430742960423231, + 0.006797171197831631, + 0.01665392331779003, + 0.014299258589744568, + 0.0074794963002204895, + 0.020196665078401566, + 0.017933247610926628, + 0.026916837319731712, + 0.004264045041054487, + 0.005094676744192839, + 0.012044585309922695, + 0.01318872906267643, + 0.007050361949950457, + 0.0064960685558617115, + 0.043633781373500824, + 0.042805057018995285, + 0.0033716419711709023, + 0.0053996434435248375, + 0.007375972345471382, + 0.008793942630290985, + 0.0037239270750433207, + 0.005475113168358803, + 0.019972922280430794, + 0.019528992474079132, + 0.006182112265378237, + 0.011806107126176357, + 0.011019062250852585, + 0.021150335669517517, + 0.008481097407639027, + 0.015373648144304752, + 0.018292797729372978, + 0.01723124459385872, + 0.0016789482906460762, + 0.002483043121173978, + 0.004568883683532476, + 0.00569760799407959, + 0.002270285040140152, + 0.0027740630321204662, + 0.0038436446338891983, + 0.004306994378566742, + 0.004954928532242775, + 0.007082142401486635, + 0.008559097535908222, + 0.009671555832028389, + 0.005444362293928862, + 0.009004282765090466, + 0.018396778032183647, + 0.011262182146310806, + 0.004491676110774279, + 0.005359978415071964, + 0.00894603319466114, + 0.011522234417498112, + 0.006740014534443617, + 0.010191359557211399, + 0.012957313098013401, + 0.009327348321676254, + 0.007466227747499943, + 0.008890558034181595, + 0.00829766970127821, + 0.007905114442110062, + 0.006273545324802399, + 0.015688998624682426, + 0.008286929689347744, + 0.00738955195993185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.004633641801774502 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605839654803276 + ], + "opposite_logprob_tokens": [ + -4.141058444976807 + ], + "logprob_reward": -0.01605839654803276, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605839654803276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07082287967205048, + 0.0410798080265522, + 0.08524134755134583, + 0.04493001848459244, + 0.35649076104164124, + 0.041468337178230286, + 0.043244726955890656, + 0.12584172189235687, + 0.12038030475378036, + 0.07050006836652756 + ], + "credit_map": [ + 0.00563883176073432, + 0.004167546518146992, + 0.01699415035545826, + 0.01112227700650692, + 0.008520067669451237, + 0.004970963578671217, + 0.0110813332721591, + 0.008327708579599857, + 0.002234937623143196, + 0.003918815869837999, + 0.007549089379608631, + 0.006989354733377695, + 0.003962269052863121, + 0.0032872420269995928, + 0.005899961572140455, + 0.007238135207444429, + 0.0059714955277740955, + 0.009114457294344902, + 0.014016181230545044, + 0.014217227697372437, + 0.0077579510398209095, + 0.010145165957510471, + 0.010563000105321407, + 0.01345586683601141, + 0.004410351160913706, + 0.0030701879877597094, + 0.005498781334608793, + 0.00862833485007286, + 0.004217548295855522, + 0.0055334060452878475, + 0.007331630680710077, + 0.006239775568246841, + 0.011023100465536118, + 0.014002311043441296, + 0.024352215230464935, + 0.036064282059669495, + 0.009465119801461697, + 0.019935425370931625, + 0.1162639707326889, + 0.12538431584835052, + 0.003116452135145664, + 0.0052849543280899525, + 0.005834810435771942, + 0.007129110395908356, + 0.003658290021121502, + 0.005073306150734425, + 0.0057900650426745415, + 0.0055813416838645935, + 0.003692679340019822, + 0.008274875581264496, + 0.006370567716658115, + 0.007155957166105509, + 0.004004696384072304, + 0.0047143734991550446, + 0.003990439232438803, + 0.005041133612394333, + 0.005740209948271513, + 0.013451213017106056, + 0.007314580492675304, + 0.012543418444693089, + 0.006843250710517168, + 0.009597184136509895, + 0.04066375270485878, + 0.029688112437725067, + 0.01024143397808075, + 0.01820310205221176, + 0.010126572102308273, + 0.026132330298423767, + 0.007943586446344852, + 0.011447684839367867, + 0.022707240656018257, + 0.013578350655734539, + 0.006533520761877298, + 0.009426889941096306, + 0.0068649291060864925, + 0.008088788948953152, + 0.006915519945323467, + 0.015965107828378677, + 0.010327635332942009, + 0.006377670913934708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.01605839654803276 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22553004324436188 + ], + "opposite_logprob_tokens": [ + -1.6005300283432007 + ], + "logprob_reward": -0.22553004324436188, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22553004324436188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2774715721607208, + 0.07476019114255905, + 0.102777861058712, + 0.05522337928414345, + 0.13769836723804474, + 0.04359288513660431, + 0.026589669287204742, + 0.0747634544968605, + 0.06912221014499664, + 0.13800033926963806 + ], + "credit_map": [ + 0.008848489262163639, + 0.004923160187900066, + 0.02376129850745201, + 0.0245046429336071, + 0.009443414397537708, + 0.007470776792615652, + 0.1097664088010788, + 0.08875339478254318, + 0.00406987126916647, + 0.0023351646959781647, + 0.02239120751619339, + 0.010867459699511528, + 0.004157537128776312, + 0.004596362356096506, + 0.007770737633109093, + 0.018571846187114716, + 0.005596562288701534, + 0.006808258127421141, + 0.010737767443060875, + 0.009766551665961742, + 0.005204478278756142, + 0.020567163825035095, + 0.015515570528805256, + 0.028581514954566956, + 0.0019917828030884266, + 0.0021448545157909393, + 0.008975444361567497, + 0.008567728102207184, + 0.002361603546887636, + 0.0022082903888076544, + 0.014282740652561188, + 0.014690936543047428, + 0.007981552742421627, + 0.010588046163320541, + 0.012924669310450554, + 0.020934389904141426, + 0.006622659042477608, + 0.017224371433258057, + 0.028521548956632614, + 0.03290112689137459, + 0.0027988923247903585, + 0.005346613470464945, + 0.005874420050531626, + 0.007961196824908257, + 0.003591604996472597, + 0.006704953033477068, + 0.005302507430315018, + 0.00601269630715251, + 0.0024814349599182606, + 0.006364054512232542, + 0.0035186621826142073, + 0.0032013983000069857, + 0.002434355905279517, + 0.002897117752581835, + 0.0029337145388126373, + 0.0027589306700974703, + 0.0033493302762508392, + 0.005352141335606575, + 0.0049817669205367565, + 0.01290519256144762, + 0.0033139095176011324, + 0.008730372413992882, + 0.019267333671450615, + 0.016863416880369186, + 0.00331379403360188, + 0.004112030379474163, + 0.0073957620188593864, + 0.017199385911226273, + 0.003817399265244603, + 0.005946081597357988, + 0.018082361668348312, + 0.009255393408238888, + 0.016749359667301178, + 0.016340864822268486, + 0.011108115315437317, + 0.014811835251748562, + 0.01214530784636736, + 0.03763387352228165, + 0.016814179718494415, + 0.012396806851029396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.22553004324436188 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, high-angle drone shot unfolds beneath a cloudless, azure sky, capturing a seemingly boundless expanse of desert. The camera then executes a swift and graceful swoop, rapidly descending to the ground, its movement fluid and purposeful. As it skims mere inches above the surface, it carefully reveals intricate, wind-sculpted ripples in the sand, their delicate patterns catching the brilliant light. The camera glides past the occasional, sparse stand of saguaro cacti, their weathered, dark green, ribbed forms casting long, stark shadows across the golden ground. Without pausing, the camera steadily and majestically ascends, pulling back to a breathtaking aerial view that re-emphasizes the sheer, overwhelming scale of the arid landscape. The entire scene is bathed in the warm, brilliant glow of the afternoon sun, igniting the rolling dunes in rich hues of deep orange and rich ochre, stretching endlessly to a shimmering, heat-hazed horizon, creating a powerful atmosphere of tranquility, majesty, and beautiful desolation.", + "seed": 26237, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_15/index_6828_seed_26237_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008246183092705906, + "vlm_logprob": -0.008246183092705906, + "vlm_yes_no_margin": 5.525, + "vlm_token_credit": -0.008246183092705906, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009479796281084418 + ], + "opposite_logprob_tokens": [ + -8.375947952270508 + ], + "logprob_reward": -0.0009479796281084418, + "yes_no_margin_reward": 8.375, + "token_credit_reward": -0.0009479796281084418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30310675501823425, + 0.02180655114352703, + 0.08091454952955246, + 0.11874089390039444, + 0.04671792685985565, + 0.10451271384954453, + 0.045868922024965286, + 0.03178369998931885, + 0.06910499930381775, + 0.17744304239749908 + ], + "credit_map": [ + 0.013407891616225243, + 0.010079349391162395, + 0.056801967322826385, + 0.0801527202129364, + 0.04237615317106247, + 0.05949064716696739, + 0.02578708715736866, + 0.015010911971330643, + 0.004055118188261986, + 0.00179259420838207, + 0.002709433436393738, + 0.0023978871759027243, + 0.0019890270195901394, + 0.002750978572294116, + 0.002715454204007983, + 0.003396057290956378, + 0.007797892205417156, + 0.007569810841232538, + 0.013571158051490784, + 0.015143219381570816, + 0.006893326994031668, + 0.006191084161400795, + 0.014194830320775509, + 0.009553213603794575, + 0.008272656239569187, + 0.007299864664673805, + 0.012843042612075806, + 0.01082353200763464, + 0.01037202961742878, + 0.012254961766302586, + 0.029478244483470917, + 0.027396555989980698, + 0.0035096341744065285, + 0.004991698544472456, + 0.009866127744317055, + 0.0057539851404726505, + 0.0040664514526724815, + 0.003968800883740187, + 0.008740042336285114, + 0.005821186117827892, + 0.013013389892876148, + 0.014159733429551125, + 0.014964998699724674, + 0.013421228155493736, + 0.013432368636131287, + 0.0118450578302145, + 0.010937657207250595, + 0.012738276273012161, + 0.004255613312125206, + 0.00726737268269062, + 0.0065354956313967705, + 0.0041249096393585205, + 0.00760379433631897, + 0.005285749211907387, + 0.006672231946140528, + 0.0041237506084144115, + 0.0032652306836098433, + 0.0027412036433815956, + 0.004965688101947308, + 0.004531210288405418, + 0.005651186220347881, + 0.003190208924934268, + 0.004334358964115381, + 0.0031046103686094284, + 0.006928035989403725, + 0.00817678589373827, + 0.010220643132925034, + 0.008538194000720978, + 0.008634696714580059, + 0.009746553376317024, + 0.007436774671077728, + 0.00942330714315176, + 0.01965080015361309, + 0.028558997437357903, + 0.01052374579012394, + 0.01975572481751442, + 0.05541083961725235, + 0.015703769400715828, + 0.015203668735921383, + 0.012635488063097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0009479796281084418 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254394743591547 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.01254394743591547, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254394743591547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09454318135976791, + 0.15291312336921692, + 0.10663175582885742, + 0.1676722764968872, + 0.04597735404968262, + 0.04915253072977066, + 0.07143422216176987, + 0.09939894825220108, + 0.15491291880607605, + 0.05736377462744713 + ], + "credit_map": [ + 0.007787997834384441, + 0.005635988432914019, + 0.01754806935787201, + 0.022735152393579483, + 0.010258976370096207, + 0.014680372551083565, + 0.00956881046295166, + 0.006327795330435038, + 0.019823458045721054, + 0.011041397228837013, + 0.02164597064256668, + 0.017077861353754997, + 0.012378965504467487, + 0.01843617670238018, + 0.021571002900600433, + 0.030938265845179558, + 0.008711839094758034, + 0.010803830809891224, + 0.017159143462777138, + 0.01852741278707981, + 0.006956619676202536, + 0.006881362292915583, + 0.022471727803349495, + 0.015119809657335281, + 0.01179021131247282, + 0.01486915536224842, + 0.018782777711749077, + 0.034006066620349884, + 0.012795576825737953, + 0.018247004598379135, + 0.0345936119556427, + 0.022587845101952553, + 0.003266029292717576, + 0.005097821820527315, + 0.006439788267016411, + 0.004607861861586571, + 0.005716852378100157, + 0.0036571503151208162, + 0.010113522410392761, + 0.007078323978930712, + 0.00777065334841609, + 0.006295381113886833, + 0.006024899426847696, + 0.006671696435660124, + 0.006746240891516209, + 0.004453438799828291, + 0.005599277559667826, + 0.005590939428657293, + 0.005061387084424496, + 0.009894120506942272, + 0.009018892422318459, + 0.007624360267072916, + 0.012653077952563763, + 0.00890054740011692, + 0.011845262721180916, + 0.006436564493924379, + 0.0095026521012187, + 0.01298667024821043, + 0.011564248241484165, + 0.016973938792943954, + 0.016409536823630333, + 0.010020985268056393, + 0.01179092563688755, + 0.010149979963898659, + 0.015062910504639149, + 0.022139471024274826, + 0.015213831327855587, + 0.01294054463505745, + 0.0315728522837162, + 0.020755218341946602, + 0.016580335795879364, + 0.020647719502449036, + 0.005556107964366674, + 0.0049183983355760574, + 0.006396197248250246, + 0.012566706165671349, + 0.006125731859356165, + 0.0055406298488378525, + 0.01000493299216032, + 0.0062550646252930164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.01254394743591547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028194934129714966 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.0028194934129714966, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028194934129714966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1794753074645996, + 0.0573095865547657, + 0.06364671885967255, + 0.3401360809803009, + 0.019434373825788498, + 0.01710996776819229, + 0.04323650524020195, + 0.198414146900177, + 0.036293067038059235, + 0.044944241642951965 + ], + "credit_map": [ + 0.007807785179466009, + 0.005836222320795059, + 0.026976773515343666, + 0.04178590327501297, + 0.03420022502541542, + 0.043418217450380325, + 0.012667136266827583, + 0.0067830332554876804, + 0.00580811221152544, + 0.0036130687221884727, + 0.007052225060760975, + 0.0063965292647480965, + 0.004921036306768656, + 0.007960982620716095, + 0.008658083155751228, + 0.0128995506092906, + 0.010003664530813694, + 0.004744356498122215, + 0.009422062896192074, + 0.00881139375269413, + 0.005376110319048166, + 0.00630258908495307, + 0.010056489147245884, + 0.008930052630603313, + 0.018323367461562157, + 0.023715520277619362, + 0.025066664442420006, + 0.04898766055703163, + 0.03320407494902611, + 0.057559799402952194, + 0.051045361906290054, + 0.08223362267017365, + 0.0013587776338681579, + 0.0017471517203375697, + 0.0036187132354825735, + 0.0017554871737957, + 0.001727556111291051, + 0.0016221198020502925, + 0.0051186298951506615, + 0.0024859379045665264, + 0.0014647255884483457, + 0.002369558671489358, + 0.002187692327424884, + 0.002139263553544879, + 0.0024256252218037844, + 0.002004543086513877, + 0.0024034224916249514, + 0.0021151357796043158, + 0.00514404708519578, + 0.006067593581974506, + 0.009268615394830704, + 0.004259714391082525, + 0.004242700524628162, + 0.003644534619525075, + 0.005651226732879877, + 0.004958076402544975, + 0.01708192378282547, + 0.023220915347337723, + 0.023779403418302536, + 0.03720328211784363, + 0.02543904446065426, + 0.021674655377864838, + 0.025438541546463966, + 0.02457638457417488, + 0.002699201228097081, + 0.0033912884537130594, + 0.004256854299455881, + 0.008009853772819042, + 0.0028775259852409363, + 0.003441255073994398, + 0.006138064432889223, + 0.005479021463543177, + 0.005235033109784126, + 0.005907771643251181, + 0.004830230958759785, + 0.004525796044617891, + 0.0055315312929451466, + 0.006095831748098135, + 0.007330452557653189, + 0.0054875947535037994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.0028194934129714966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734531372785568 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734531372785568, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734531372785568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13214175403118134, + 0.03119521215558052, + 0.04285373166203499, + 0.10985679179430008, + 0.13315874338150024, + 0.08502176403999329, + 0.08213671296834946, + 0.1354968100786209, + 0.1053902730345726, + 0.14274829626083374 + ], + "credit_map": [ + 0.006317611783742905, + 0.005158104468137026, + 0.024147028103470802, + 0.03551614657044411, + 0.017609581351280212, + 0.02495773322880268, + 0.010936680249869823, + 0.007498840801417828, + 0.003705530194565654, + 0.002000353066250682, + 0.004212281201034784, + 0.00334706949070096, + 0.0026625962927937508, + 0.00409179599955678, + 0.004484805278480053, + 0.0066907769069075584, + 0.005305575672537088, + 0.0040426249615848064, + 0.0062740882858633995, + 0.00802824180573225, + 0.0019764513708651066, + 0.002220747759565711, + 0.007225325331091881, + 0.007780665531754494, + 0.0065112244337797165, + 0.008420080877840519, + 0.012233023531734943, + 0.014775728806853294, + 0.006369469221681356, + 0.006416741292923689, + 0.03293716162443161, + 0.022193342447280884, + 0.010042671114206314, + 0.009959463961422443, + 0.032478634268045425, + 0.01353641226887703, + 0.009224636480212212, + 0.016004817560315132, + 0.02397526055574417, + 0.01793682761490345, + 0.011572149582207203, + 0.00781458243727684, + 0.018233804032206535, + 0.011503927409648895, + 0.008775351569056511, + 0.009976963512599468, + 0.007390476763248444, + 0.009754491038620472, + 0.0059380363672971725, + 0.0104429442435503, + 0.01105946209281683, + 0.007386456243693829, + 0.021689504384994507, + 0.008642720989882946, + 0.011650342494249344, + 0.005327223800122738, + 0.015333465300500393, + 0.01669444888830185, + 0.013431585393846035, + 0.020747678354382515, + 0.018563905730843544, + 0.025334171950817108, + 0.014225535094738007, + 0.01116598304361105, + 0.008784930221736431, + 0.015101372264325619, + 0.013064556755125523, + 0.009291645139455795, + 0.01908055879175663, + 0.012836207635700703, + 0.012867570854723454, + 0.014363409020006657, + 0.011912058107554913, + 0.027649253606796265, + 0.010911090299487114, + 0.016128217801451683, + 0.030483217909932137, + 0.015099702402949333, + 0.01638195849955082, + 0.014182777144014835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.006734531372785568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818496361374855 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.01818496361374855, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818496361374855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17463985085487366, + 0.1341542899608612, + 0.044837966561317444, + 0.14328071475028992, + 0.0474638007581234, + 0.02312537655234337, + 0.14866408705711365, + 0.169844850897789, + 0.058205582201480865, + 0.0557834766805172 + ], + "credit_map": [ + 0.006455868482589722, + 0.005380748771131039, + 0.029749903827905655, + 0.045601196587085724, + 0.029075590893626213, + 0.03795173764228821, + 0.013361119665205479, + 0.007063689175993204, + 0.017947165295481682, + 0.008685903623700142, + 0.01817326620221138, + 0.014718472957611084, + 0.010579373687505722, + 0.01640673354268074, + 0.019949473440647125, + 0.027693910524249077, + 0.005063462071120739, + 0.005572246387600899, + 0.007107092533260584, + 0.009280109778046608, + 0.002439711708575487, + 0.0023968545719981194, + 0.007091233041137457, + 0.005887261591851711, + 0.007241284940391779, + 0.00889857579022646, + 0.009769708849489689, + 0.0129717281088233, + 0.006744931451976299, + 0.011243275366723537, + 0.04416908323764801, + 0.0422421358525753, + 0.0039055608212947845, + 0.004088029731065035, + 0.008710167370736599, + 0.004964048974215984, + 0.0039506945759058, + 0.007523275911808014, + 0.00693523557856679, + 0.007386791054159403, + 0.0030642400961369276, + 0.003198816906660795, + 0.003009482752531767, + 0.004317573271691799, + 0.001934381783939898, + 0.002143997699022293, + 0.002615791978314519, + 0.0028410933446139097, + 0.01423050556331873, + 0.029393717646598816, + 0.020589057356119156, + 0.012915742583572865, + 0.015317538753151894, + 0.027747998014092445, + 0.014181253500282764, + 0.014288277365267277, + 0.014204325154423714, + 0.018527915701270103, + 0.01970815286040306, + 0.02849595434963703, + 0.02069985307753086, + 0.025127161294221878, + 0.02493394911289215, + 0.0181475467979908, + 0.005196897778660059, + 0.008313674479722977, + 0.00820855051279068, + 0.004698434378951788, + 0.010440647602081299, + 0.008079782128334045, + 0.005913048982620239, + 0.007354548666626215, + 0.006093205418437719, + 0.006606992334127426, + 0.006015794817358255, + 0.011204445734620094, + 0.006483615376055241, + 0.0036326751578599215, + 0.009475366212427616, + 0.006271385122090578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.01818496361374855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, high-angle drone shot unfolds beneath a cloudless, azure sky, capturing a seemingly boundless expanse of desert. The camera then executes a swift and graceful swoop, rapidly descending to the ground, its movement fluid and purposeful. As it skims mere inches above the surface, it carefully reveals intricate, wind-sculpted ripples in the sand, their delicate patterns catching the brilliant light. The camera glides past the occasional, sparse stand of saguaro cacti, their weathered, dark green, ribbed forms casting long, stark shadows across the golden ground. Without pausing, the camera steadily and majestically ascends, pulling back to a breathtaking aerial view that re-emphasizes the sheer, overwhelming scale of the arid landscape. The entire scene is bathed in the warm, brilliant glow of the afternoon sun, igniting the rolling dunes in rich hues of deep orange and rich ochre, stretching endlessly to a shimmering, heat-hazed horizon, creating a powerful atmosphere of tranquility, majesty, and beautiful desolation.", + "seed": 26238, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_15/index_6828_seed_26238_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014587492612190545, + "vlm_logprob": -0.014587492612190545, + "vlm_yes_no_margin": 5.049999952316284, + "vlm_token_credit": -0.014587492612190545, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0018938715802505612 + ], + "opposite_logprob_tokens": [ + -7.876893997192383 + ], + "logprob_reward": -0.0018938715802505612, + "yes_no_margin_reward": 7.875, + "token_credit_reward": -0.0018938715802505612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11444279551506042, + 0.11756835132837296, + 0.06989037245512009, + 0.12021004408597946, + 0.13178479671478271, + 0.11824464052915573, + 0.17003051936626434, + 0.032860167324543, + 0.07247094064950943, + 0.05249738693237305 + ], + "credit_map": [ + 0.006762878969311714, + 0.0065496014431118965, + 0.025659291073679924, + 0.022236136719584465, + 0.014707583002746105, + 0.009338832460343838, + 0.016241205856204033, + 0.012947270646691322, + 0.005843510385602713, + 0.008067268878221512, + 0.01485731266438961, + 0.02052062377333641, + 0.021845143288373947, + 0.009329130873084068, + 0.019078893586993217, + 0.01802646555006504, + 0.007642928510904312, + 0.005117980297654867, + 0.012288370169699192, + 0.015432054176926613, + 0.004723162390291691, + 0.005651375744491816, + 0.009904437698423862, + 0.00913006067276001, + 0.006751888431608677, + 0.013250760734081268, + 0.025884849950671196, + 0.025313297286629677, + 0.010298279114067554, + 0.009200305677950382, + 0.01771162636578083, + 0.011799031868577003, + 0.004690997302532196, + 0.005787775386124849, + 0.008084787987172604, + 0.012726386077702045, + 0.004757656250149012, + 0.00794715341180563, + 0.047277290374040604, + 0.04051274433732033, + 0.007714372128248215, + 0.013670939952135086, + 0.014328224584460258, + 0.015240686945617199, + 0.011327284388244152, + 0.0154185201972723, + 0.017603177577257156, + 0.022941438481211662, + 0.006224201526492834, + 0.011164921335875988, + 0.02168687805533409, + 0.032892029732465744, + 0.011426880955696106, + 0.022214828059077263, + 0.03428095951676369, + 0.03013981133699417, + 0.0016646322328597307, + 0.0032852652948349714, + 0.004454135894775391, + 0.003813409013673663, + 0.0032855451572686434, + 0.006009639706462622, + 0.00462881475687027, + 0.005718724336475134, + 0.006818845868110657, + 0.007187791634351015, + 0.016526687890291214, + 0.009727226570248604, + 0.005553706549108028, + 0.00833899062126875, + 0.013748270459473133, + 0.004569417797029018, + 0.005594537127763033, + 0.011878917925059795, + 0.009208997711539268, + 0.005817814730107784, + 0.005509954411536455, + 0.0064283874817192554, + 0.004217791836708784, + 0.0038409866392612457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.0018938715802505612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584993064403534 + ], + "opposite_logprob_tokens": [ + -3.895585060119629 + ], + "logprob_reward": -0.020584993064403534, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584993064403534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22278562188148499, + 0.15316715836524963, + 0.056469712406396866, + 0.09811242669820786, + 0.1636217087507248, + 0.07446808367967606, + 0.1238371804356575, + 0.020073536783456802, + 0.05641641840338707, + 0.0310481246560812 + ], + "credit_map": [ + 0.019698480144143105, + 0.03824334591627121, + 0.014364766888320446, + 0.03067202866077423, + 0.05026139318943024, + 0.022826485335826874, + 0.02533455193042755, + 0.021384581923484802, + 0.008579110726714134, + 0.01262219063937664, + 0.009890330955386162, + 0.012150514870882034, + 0.049021460115909576, + 0.019728699699044228, + 0.013505878858268261, + 0.027668965980410576, + 0.003910493105649948, + 0.0045511736534535885, + 0.01079498790204525, + 0.012511654756963253, + 0.006636901758611202, + 0.0046398150734603405, + 0.006956410128623247, + 0.006468274164944887, + 0.004942332860082388, + 0.008447394706308842, + 0.016367092728614807, + 0.026946505531668663, + 0.007839657366275787, + 0.006118703167885542, + 0.016558192670345306, + 0.010892539285123348, + 0.002800178714096546, + 0.004294209182262421, + 0.005974754225462675, + 0.012784602120518684, + 0.006788000464439392, + 0.011136429384350777, + 0.06422196328639984, + 0.055621564388275146, + 0.003372824750840664, + 0.008055432699620724, + 0.006990890484303236, + 0.016608476638793945, + 0.004882750101387501, + 0.0060072545893490314, + 0.01770656928420067, + 0.01084388792514801, + 0.006105019710958004, + 0.005643942393362522, + 0.012911603786051273, + 0.025434907525777817, + 0.005948877893388271, + 0.020300090312957764, + 0.02600470371544361, + 0.021488038823008537, + 0.0007209783070720732, + 0.0017992182401940227, + 0.0031024569179862738, + 0.0029329685494303703, + 0.0017724072095006704, + 0.003136012703180313, + 0.003159242682158947, + 0.0034502511844038963, + 0.004985579289495945, + 0.0058672428131103516, + 0.01477936003357172, + 0.007435674779117107, + 0.003794462885707617, + 0.0037487135268747807, + 0.012742241844534874, + 0.00306314486078918, + 0.001992132281884551, + 0.003148844465613365, + 0.00415711710229516, + 0.008859528228640556, + 0.006638106890022755, + 0.0018134835409000516, + 0.002906892914324999, + 0.0015320192323997617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.020584993064403534 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024874242953956127 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024874242953956127, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024874242953956127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2881007790565491, + 0.27233895659446716, + 0.0618608184158802, + 0.05221512168645859, + 0.10221482068300247, + 0.06925494223833084, + 0.046170711517333984, + 0.03335646912455559, + 0.03467048332095146, + 0.03981691226363182 + ], + "credit_map": [ + 0.025138046592473984, + 0.044787850230932236, + 0.026933791115880013, + 0.04493032395839691, + 0.06282278150320053, + 0.029107365757226944, + 0.03084825538098812, + 0.02353234775364399, + 0.020429983735084534, + 0.03094875067472458, + 0.038421209901571274, + 0.05851849913597107, + 0.01923445053398609, + 0.013361040502786636, + 0.04865628480911255, + 0.04276873916387558, + 0.004284322261810303, + 0.00409564608708024, + 0.012896467931568623, + 0.015078091062605381, + 0.006389157846570015, + 0.0030765836127102375, + 0.008239834569394588, + 0.007800715509802103, + 0.004038022831082344, + 0.006546071730554104, + 0.0070541962049901485, + 0.009493461810052395, + 0.004031819757074118, + 0.006177485454827547, + 0.007126749027520418, + 0.007747315801680088, + 0.002840717090293765, + 0.003546706400811672, + 0.004832386039197445, + 0.009807433933019638, + 0.005011504981666803, + 0.006684510037302971, + 0.03688321262598038, + 0.032608356326818466, + 0.003998264670372009, + 0.00850906502455473, + 0.010145319625735283, + 0.00893919263035059, + 0.006322972476482391, + 0.008868048898875713, + 0.01028792466968298, + 0.012184147723019123, + 0.0027314452454447746, + 0.00328555959276855, + 0.006120841018855572, + 0.007737048901617527, + 0.00334221706725657, + 0.004598172847181559, + 0.008504447527229786, + 0.009850977919995785, + 0.0017067574663087726, + 0.0024332620669156313, + 0.005918366368860006, + 0.00677706953138113, + 0.0025020933244377375, + 0.003807242028415203, + 0.005919269286096096, + 0.004292412661015987, + 0.002942742547020316, + 0.0035387901589274406, + 0.006268270779401064, + 0.005633587017655373, + 0.004048102535307407, + 0.0021798231173306704, + 0.008151127956807613, + 0.0019080420024693012, + 0.0032278948929160833, + 0.005875943694263697, + 0.008097575046122074, + 0.007054904010146856, + 0.005086956545710564, + 0.002923489548265934, + 0.005311237182468176, + 0.0022389106452465057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0024874242953956127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178174272179604 + ], + "opposite_logprob_tokens": [ + -4.0181779861450195 + ], + "logprob_reward": -0.018178174272179604, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018178174272179604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26420214772224426, + 0.06644784659147263, + 0.05863644555211067, + 0.07949619740247726, + 0.05140188708901405, + 0.19114278256893158, + 0.04270463064312935, + 0.0842016413807869, + 0.10021410137414932, + 0.06155235692858696 + ], + "credit_map": [ + 0.019911766052246094, + 0.04287852346897125, + 0.016457535326480865, + 0.03189831227064133, + 0.0718509778380394, + 0.032575853168964386, + 0.0282182265073061, + 0.020410945639014244, + 0.0037375292740762234, + 0.005611628759652376, + 0.006338944658637047, + 0.008183668367564678, + 0.018511289730668068, + 0.0074968719854950905, + 0.005901476368308067, + 0.010666437447071075, + 0.005357010755687952, + 0.0030970473308116198, + 0.013904791325330734, + 0.0161929689347744, + 0.0030354459304362535, + 0.004023561254143715, + 0.006610450334846973, + 0.0064151654951274395, + 0.004627794027328491, + 0.0076277959160506725, + 0.013009410351514816, + 0.016887811943888664, + 0.007680506445467472, + 0.008037780411541462, + 0.011751953512430191, + 0.009873149916529655, + 0.002078315243124962, + 0.0023580659180879593, + 0.003337166039273143, + 0.004943531937897205, + 0.0023968438617885113, + 0.0028541216161102057, + 0.01815272867679596, + 0.015281113795936108, + 0.006076921243220568, + 0.01573488488793373, + 0.01576272025704384, + 0.05137891694903374, + 0.010150298476219177, + 0.01037538517266512, + 0.05594674497842789, + 0.02571691758930683, + 0.002179198432713747, + 0.003476545447483659, + 0.00440585520118475, + 0.010121970437467098, + 0.0035670313518494368, + 0.004692126531153917, + 0.007537885569036007, + 0.006724016275256872, + 0.0037535284645855427, + 0.007398249581456184, + 0.012828663922846317, + 0.010590783320367336, + 0.006095510441809893, + 0.01616034097969532, + 0.0125600416213274, + 0.014814525842666626, + 0.009073489345610142, + 0.011342056095600128, + 0.012480120174586773, + 0.02432139590382576, + 0.009602505713701248, + 0.007999090477824211, + 0.019078295677900314, + 0.0063171470537781715, + 0.0060331085696816444, + 0.0100729800760746, + 0.009724250994622707, + 0.005935397464782, + 0.01256945077329874, + 0.006831466220319271, + 0.0066178846172988415, + 0.003767817746847868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.018178174272179604 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979299984872341 + ], + "opposite_logprob_tokens": [ + -3.5297930240631104 + ], + "logprob_reward": -0.02979299984872341, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979299984872341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2261378914117813, + 0.24618828296661377, + 0.0614580474793911, + 0.07817579060792923, + 0.09400651603937149, + 0.16492187976837158, + 0.05979292467236519, + 0.016075164079666138, + 0.013271842151880264, + 0.03997164964675903 + ], + "credit_map": [ + 0.01908928155899048, + 0.03908507153391838, + 0.01555513683706522, + 0.02753433957695961, + 0.05647949129343033, + 0.025793645530939102, + 0.02479526214301586, + 0.017805660143494606, + 0.017311761155724525, + 0.02461308240890503, + 0.017396820709109306, + 0.022694258019328117, + 0.07742566615343094, + 0.032264530658721924, + 0.020093008875846863, + 0.03438914939761162, + 0.004763088654726744, + 0.004736427683383226, + 0.010886677540838718, + 0.013818084262311459, + 0.008863036520779133, + 0.0064160265028476715, + 0.006228829734027386, + 0.0057458714582026005, + 0.006378428544849157, + 0.007716177962720394, + 0.014918360859155655, + 0.015685580670833588, + 0.007128180470317602, + 0.004210146609693766, + 0.012151459231972694, + 0.009987463243305683, + 0.002637110650539398, + 0.002643005922436714, + 0.0036752130836248398, + 0.008773905225098133, + 0.005223364569246769, + 0.0071177612990140915, + 0.035138484090566635, + 0.02879767119884491, + 0.005931778345257044, + 0.013120763935148716, + 0.01160765253007412, + 0.052469927817583084, + 0.009622723795473576, + 0.0074398512952029705, + 0.0497211292386055, + 0.0150080481544137, + 0.0030510465148836374, + 0.005379681009799242, + 0.0061352793127298355, + 0.011228925548493862, + 0.005093997810035944, + 0.007900011725723743, + 0.011327989399433136, + 0.009675992652773857, + 0.0007663731812499464, + 0.0013124893885105848, + 0.002827160758897662, + 0.002804249757900834, + 0.0011268026428297162, + 0.0016683000139892101, + 0.0033136666752398014, + 0.0022561217192560434, + 0.0017611104995012283, + 0.0014021528186276555, + 0.0020228573121130466, + 0.0017897840589284897, + 0.0010702500585466623, + 0.0015409798361361027, + 0.0023382436484098434, + 0.0013464646181091666, + 0.0032900862861424685, + 0.005872843321412802, + 0.007682010531425476, + 0.0047690290957689285, + 0.004401232115924358, + 0.00562572106719017, + 0.004963801242411137, + 0.0033669269178062677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.02979299984872341 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, high-angle drone shot unfolds beneath a cloudless, azure sky, capturing a seemingly boundless expanse of desert. The camera then executes a swift and graceful swoop, rapidly descending to the ground, its movement fluid and purposeful. As it skims mere inches above the surface, it carefully reveals intricate, wind-sculpted ripples in the sand, their delicate patterns catching the brilliant light. The camera glides past the occasional, sparse stand of saguaro cacti, their weathered, dark green, ribbed forms casting long, stark shadows across the golden ground. Without pausing, the camera steadily and majestically ascends, pulling back to a breathtaking aerial view that re-emphasizes the sheer, overwhelming scale of the arid landscape. The entire scene is bathed in the warm, brilliant glow of the afternoon sun, igniting the rolling dunes in rich hues of deep orange and rich ochre, stretching endlessly to a shimmering, heat-hazed horizon, creating a powerful atmosphere of tranquility, majesty, and beautiful desolation.", + "seed": 26239, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_15/index_6828_seed_26239_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010655152820982039, + "vlm_logprob": -0.010655152820982039, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.010655152820982039, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011235122801735997 + ], + "opposite_logprob_tokens": [ + -7.626123428344727 + ], + "logprob_reward": -0.0011235122801735997, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0011235122801735997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07777988165616989, + 0.11694779247045517, + 0.12713667750358582, + 0.21791639924049377, + 0.0758051723241806, + 0.025057896971702576, + 0.036195866763591766, + 0.09800436347723007, + 0.1374308317899704, + 0.08772511035203934 + ], + "credit_map": [ + 0.009167062118649483, + 0.0071917930617928505, + 0.01201374176889658, + 0.013702689670026302, + 0.006974787916988134, + 0.007422411814332008, + 0.012183543294668198, + 0.00912385992705822, + 0.010139926336705685, + 0.014080609194934368, + 0.012812215834856033, + 0.03637650981545448, + 0.005729969125241041, + 0.005694023799151182, + 0.020355554297566414, + 0.01175900548696518, + 0.010984301567077637, + 0.014717597514390945, + 0.01715606264770031, + 0.01710987649857998, + 0.013975251466035843, + 0.014874951913952827, + 0.021349120885133743, + 0.01696952059864998, + 0.017319263890385628, + 0.011649155989289284, + 0.0759667232632637, + 0.07780037820339203, + 0.008427692577242851, + 0.005851855967193842, + 0.014272286556661129, + 0.0066290684044361115, + 0.008578994311392307, + 0.007658917456865311, + 0.009550950489938259, + 0.013424159958958626, + 0.006500000134110451, + 0.006921176332980394, + 0.013999458402395248, + 0.009171522222459316, + 0.002337430836632848, + 0.004178404808044434, + 0.003968335222452879, + 0.003902279306203127, + 0.0016263340367004275, + 0.0037932279519736767, + 0.003578483359888196, + 0.0016734059900045395, + 0.0038357418961822987, + 0.003957390319555998, + 0.0057047209702432156, + 0.0066142757423222065, + 0.004955837037414312, + 0.0026100827381014824, + 0.0027484362944960594, + 0.005769387353211641, + 0.020034952089190483, + 0.011126468889415264, + 0.015218693763017654, + 0.008376567624509335, + 0.010136212222278118, + 0.010505267418920994, + 0.01167423278093338, + 0.01093197800219059, + 0.01693863607943058, + 0.015543450601398945, + 0.007543805055320263, + 0.012463854625821114, + 0.02383238449692726, + 0.03574057295918465, + 0.008668431080877781, + 0.016699710860848427, + 0.007512698415666819, + 0.010733061470091343, + 0.011737139895558357, + 0.011900648474693298, + 0.010283985175192356, + 0.00964295119047165, + 0.01104481890797615, + 0.014869811944663525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.0011235122801735997 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083347722887993 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011083347722887993, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083347722887993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0407133586704731, + 0.06034819409251213, + 0.12627825140953064, + 0.07223174721002579, + 0.11008204519748688, + 0.03014267049729824, + 0.11424118280410767, + 0.09530644863843918, + 0.10092448443174362, + 0.24973160028457642 + ], + "credit_map": [ + 0.004081047140061855, + 0.0036054791416972876, + 0.006639610510319471, + 0.007587369065731764, + 0.003900077659636736, + 0.00383737962692976, + 0.006040005944669247, + 0.005022390745580196, + 0.004323069471865892, + 0.005111933685839176, + 0.007097179535776377, + 0.019869642332196236, + 0.002725922968238592, + 0.0025312926154583693, + 0.011969073675572872, + 0.006720082834362984, + 0.006004830822348595, + 0.00788139645010233, + 0.020744046196341515, + 0.046698734164237976, + 0.007730415556579828, + 0.006217246875166893, + 0.022784994915127754, + 0.008216592483222485, + 0.011599411256611347, + 0.00814568530768156, + 0.011626929976046085, + 0.011023986153304577, + 0.009906464256346226, + 0.006085908971726894, + 0.008682209067046642, + 0.005161151755601168, + 0.018832135945558548, + 0.011435729451477528, + 0.01403531339019537, + 0.017188705503940582, + 0.007594057824462652, + 0.010288362391293049, + 0.019534654915332794, + 0.011173093691468239, + 0.004133675713092089, + 0.003416720312088728, + 0.0046033901162445545, + 0.004682632163167, + 0.002266880124807358, + 0.0034442576579749584, + 0.003650583093985915, + 0.00394453015178442, + 0.011622845195233822, + 0.011508328840136528, + 0.016709990799427032, + 0.020164277404546738, + 0.015491778030991554, + 0.013263266533613205, + 0.012515554204583168, + 0.012965145520865917, + 0.019462471827864647, + 0.009186508134007454, + 0.00976574793457985, + 0.01005763839930296, + 0.008455449715256691, + 0.014033673331141472, + 0.012164662592113018, + 0.012180290184915066, + 0.01082595158368349, + 0.009386882185935974, + 0.007020825520157814, + 0.02354861982166767, + 0.011018357239663601, + 0.015248224139213562, + 0.008437639102339745, + 0.01543798204511404, + 0.011031672358512878, + 0.013210286386311054, + 0.016740946099162102, + 0.03088965266942978, + 0.013675984926521778, + 0.009999617002904415, + 0.09017584472894669, + 0.06400760263204575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 6 + ], + "selected_reward": -0.011083347722887993 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052522532641887665 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.0052522532641887665, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052522532641887665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02888435497879982, + 0.0713222324848175, + 0.13862715661525726, + 0.08317747712135315, + 0.04674888402223587, + 0.027733298018574715, + 0.06705454736948013, + 0.04778067767620087, + 0.05512980371713638, + 0.43354153633117676 + ], + "credit_map": [ + 0.004071636591106653, + 0.003641862189397216, + 0.004090204834938049, + 0.004473243374377489, + 0.0026222695596516132, + 0.0020318294409662485, + 0.0036404835991561413, + 0.004312824923545122, + 0.005493585951626301, + 0.006588357500731945, + 0.008085406385362148, + 0.019108761101961136, + 0.0040909829549491405, + 0.00451941043138504, + 0.013060948811471462, + 0.010374778881669044, + 0.00840060319751501, + 0.009492410346865654, + 0.022633252665400505, + 0.0332813560962677, + 0.011877988465130329, + 0.010574398562312126, + 0.02116561494767666, + 0.021201536059379578, + 0.006366402842104435, + 0.005205477587878704, + 0.026495343074202538, + 0.026980172842741013, + 0.004777374677360058, + 0.0038987002335488796, + 0.005779181141406298, + 0.0036748263519257307, + 0.0062465057708323, + 0.005765380337834358, + 0.0072326031513512135, + 0.008340959437191486, + 0.004408540669828653, + 0.0033250655978918076, + 0.00644960580393672, + 0.004980224184691906, + 0.0034088643733412027, + 0.00447874516248703, + 0.004525204189121723, + 0.003995514940470457, + 0.0016035472508519888, + 0.0032098933588713408, + 0.0030578116420656443, + 0.003453716402873397, + 0.012247050181031227, + 0.005917984526604414, + 0.01045005302876234, + 0.008134866133332253, + 0.004900796338915825, + 0.010699587874114513, + 0.0046872952952980995, + 0.0100169125944376, + 0.00726788304746151, + 0.006910362746566534, + 0.007520419079810381, + 0.0046884603798389435, + 0.006149550899863243, + 0.004401321988552809, + 0.006068004295229912, + 0.004774675238877535, + 0.004863647744059563, + 0.008265781216323376, + 0.0036098030395805836, + 0.0086215203627944, + 0.00960517581552267, + 0.006909360643476248, + 0.004969857633113861, + 0.008284653536975384, + 0.01796426996588707, + 0.01454272773116827, + 0.03807777166366577, + 0.0699123963713646, + 0.014955197460949421, + 0.014043381437659264, + 0.16678179800510406, + 0.09726400673389435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.0052522532641887665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253335177898407 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.01253335177898407, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253335177898407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03344990313053131, + 0.04883065074682236, + 0.3369434177875519, + 0.14589719474315643, + 0.15532656013965607, + 0.03318439796566963, + 0.040006522089242935, + 0.06397749483585358, + 0.06764373183250427, + 0.07474011927843094 + ], + "credit_map": [ + 0.004459416959434748, + 0.0034233867190778255, + 0.005667633842676878, + 0.004934599623084068, + 0.0035568580497056246, + 0.0032365822698920965, + 0.00420781597495079, + 0.003963610157370567, + 0.004339196719229221, + 0.00413725059479475, + 0.006647750269621611, + 0.015187441371381283, + 0.0022970496211200953, + 0.0026451894082129, + 0.007349277846515179, + 0.006227497942745686, + 0.021552052348852158, + 0.03444340452551842, + 0.04968563839793205, + 0.1137964054942131, + 0.02325380966067314, + 0.023784330114722252, + 0.053579557687044144, + 0.016848208382725716, + 0.01937706768512726, + 0.017785919830203056, + 0.030938951298594475, + 0.026378175243735313, + 0.017772020772099495, + 0.010611861944198608, + 0.014993665739893913, + 0.008039534091949463, + 0.017373984679579735, + 0.019596416503190994, + 0.01954948529601097, + 0.021293340250849724, + 0.015397603623569012, + 0.022157466039061546, + 0.019795434549450874, + 0.020162831991910934, + 0.0038216093089431524, + 0.003969854675233364, + 0.005124537739902735, + 0.004846596624702215, + 0.0024010371416807175, + 0.003986261319369078, + 0.0053296322003006935, + 0.003704868257045746, + 0.005014169029891491, + 0.004584337584674358, + 0.006135764066129923, + 0.0070544951595366, + 0.004154947120696306, + 0.005130921956151724, + 0.002950109075754881, + 0.00498177669942379, + 0.013709732331335545, + 0.007644952274858952, + 0.00772112375125289, + 0.004983982536941767, + 0.007105871569365263, + 0.007410163059830666, + 0.008543194271624088, + 0.006858475040644407, + 0.008606174029409885, + 0.009382265619933605, + 0.0029754191637039185, + 0.008369272574782372, + 0.011221744120121002, + 0.015886005014181137, + 0.0030006831511855125, + 0.008202171884477139, + 0.004412224516272545, + 0.005302743054926395, + 0.005518940277397633, + 0.009862778708338737, + 0.004136031959205866, + 0.004613240249454975, + 0.02405950240790844, + 0.016834652051329613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.01253335177898407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a desert with rolling dunes under a cloudless azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle view and then swoop rapidly down to skim inches above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are wind-sculpted ripples in the sand visible as the camera skims the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are sparse saguaro cacti with dark green ribbed forms visible casting long shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera ascend and pull back to reveal the vast scale of the landscape bathed in warm afternoon light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283299058675766 + ], + "opposite_logprob_tokens": [ + -3.7732832431793213 + ], + "logprob_reward": -0.023283299058675766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283299058675766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02147863060235977, + 0.03492322191596031, + 0.1290407031774521, + 0.027583403512835503, + 0.04274144768714905, + 0.02109045349061489, + 0.03732661157846451, + 0.041469257324934006, + 0.112567238509655, + 0.5317790508270264 + ], + "credit_map": [ + 0.003283147467300296, + 0.0026235952973365784, + 0.002572648460045457, + 0.0029314951971173286, + 0.0025847568176686764, + 0.0022686731535941362, + 0.0028256415389478207, + 0.0023886726703494787, + 0.003957263194024563, + 0.004290525335818529, + 0.004306552931666374, + 0.0071770139038562775, + 0.002326602814719081, + 0.002448869403451681, + 0.005339883267879486, + 0.005076510366052389, + 0.008306392468512058, + 0.011900143697857857, + 0.01883857138454914, + 0.04180368781089783, + 0.008843712508678436, + 0.006246661301702261, + 0.02419094927608967, + 0.008910578675568104, + 0.003933903761208057, + 0.0033873978536576033, + 0.0046057733707129955, + 0.004885148257017136, + 0.003064470598474145, + 0.0019212313927710056, + 0.003261048346757889, + 0.0025244292337447405, + 0.010863707400858402, + 0.005092992912977934, + 0.005948036443442106, + 0.005601155571639538, + 0.002545958152040839, + 0.0025500908959656954, + 0.007232125382870436, + 0.0029073834884911776, + 0.0028922089841216803, + 0.0031662487890571356, + 0.003505159169435501, + 0.003280286444351077, + 0.001684150891378522, + 0.0020348539110273123, + 0.0027690764982253313, + 0.0017584683373570442, + 0.006807542406022549, + 0.0028172447346150875, + 0.005764608737081289, + 0.006510635372251272, + 0.003992332145571709, + 0.003743002424016595, + 0.003582295263186097, + 0.0041089532896876335, + 0.006463659927248955, + 0.003833471331745386, + 0.004044307861477137, + 0.00504722073674202, + 0.005174117628484964, + 0.005369782447814941, + 0.006538382265716791, + 0.004998316057026386, + 0.009801150299608707, + 0.013075224123895168, + 0.006668040528893471, + 0.02237633243203163, + 0.012972336262464523, + 0.014277298003435135, + 0.014336536638438702, + 0.019060321152210236, + 0.020506305620074272, + 0.016242099925875664, + 0.0295436792075634, + 0.07440183311700821, + 0.01950545608997345, + 0.014605785720050335, + 0.21407511830329895, + 0.1428988128900528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.023283299058675766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_2/index_tensor([104, 104])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_2/index_tensor([104, 104])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b4d48be507c9caa82e90eb4dd4892afeb923f857 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_2/index_tensor([104, 104])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot from a slightly high angle captures a captivating, sharp time-lapse of a giant Victoria Amazon water lily in full bloom. The scene begins with a spiny, dark green bud with a bronze hue emerging from the still, dark water, its surface reflecting the ever-changing ambient light. In an accelerated yet seamless sequence, the tightly curled bud begins its transformation, which seems to unfold over the course of an entire day and night. The outer petals deliberately peel away, revealing the massive, pure white inner petals. These waxy, boat-shaped petals gracefully unfold layer by layer in a rhythmic, hypnotic dance, their pristine surfaces capturing the transitional light, from the cool blue of dawn, through the bright clarity of midday, to the warm golden hues of sunset. This movement culminates when the flower fully opens, revealing its intricate golden center—a dense corolla that creates a stunning, vibrant contrast with the stark white of the blooming flower, creating a serene and awe-inspiring spectacle of natural beauty.", + "seed": 20147, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_2/index_104_seed_20147_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11519902206491679, + "vlm_logprob": -0.11519902206491679, + "vlm_yes_no_margin": 3.8499999940395355, + "vlm_token_credit": -0.11519902206491679, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07050352543592453 + ], + "opposite_logprob_tokens": [ + -2.8205034732818604 + ], + "logprob_reward": -0.07050352543592453, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07050352543592453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21728438138961792, + 0.08174946904182434, + 0.07122292369604111, + 0.06226249039173126, + 0.05824749916791916, + 0.16855686902999878, + 0.06911423802375793, + 0.1481272280216217, + 0.05516619607806206, + 0.06826864928007126 + ], + "credit_map": [ + 0.028750935569405556, + 0.015883760526776314, + 0.04466328024864197, + 0.019358471035957336, + 0.015011170879006386, + 0.01988878846168518, + 0.031414955854415894, + 0.04231303557753563, + 0.007461335975676775, + 0.008657262660562992, + 0.012917310930788517, + 0.008160884492099285, + 0.010879111476242542, + 0.007697070483118296, + 0.012156601995229721, + 0.013819894753396511, + 0.006736310198903084, + 0.008723526261746883, + 0.010410213842988014, + 0.006034248508512974, + 0.00827654916793108, + 0.007891601882874966, + 0.008518527261912823, + 0.014631951227784157, + 0.009300600737333298, + 0.0072021158412098885, + 0.007227187976241112, + 0.006027169059962034, + 0.00849170982837677, + 0.007787059992551804, + 0.005967964883893728, + 0.010258682072162628, + 0.007210349198430777, + 0.006525412667542696, + 0.0076218112371861935, + 0.004686218220740557, + 0.009537572041153908, + 0.005425407085567713, + 0.0074976528994739056, + 0.009743076749145985, + 0.008482293225824833, + 0.009543831460177898, + 0.04489251598715782, + 0.006902210880070925, + 0.014886932447552681, + 0.006560001987963915, + 0.062454912811517715, + 0.01483417209237814, + 0.011731233447790146, + 0.007788904942572117, + 0.007097212597727776, + 0.00903730746358633, + 0.008168041706085205, + 0.008662751875817776, + 0.008118636906147003, + 0.00851015280932188, + 0.01328987255692482, + 0.028108322992920876, + 0.02028026059269905, + 0.006828597281128168, + 0.04896452650427818, + 0.012307929806411266, + 0.009030905552208424, + 0.009316815063357353, + 0.004781381692737341, + 0.007295980118215084, + 0.005714356899261475, + 0.011815906502306461, + 0.007169491611421108, + 0.003434297163039446, + 0.00978122279047966, + 0.005173558834940195, + 0.012436176650226116, + 0.00798901729285717, + 0.008705331943929195, + 0.009702051058411598, + 0.005965335760265589, + 0.011242850683629513, + 0.006668636575341225, + 0.005559253040701151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.07050352543592453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638811059296131 + ], + "opposite_logprob_tokens": [ + -3.651388168334961 + ], + "logprob_reward": -0.02638811059296131, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638811059296131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19683870673179626, + 0.05969984084367752, + 0.022172031924128532, + 0.02853427082300186, + 0.05400277301669121, + 0.037971511483192444, + 0.04002610221505165, + 0.44856196641921997, + 0.07056776434183121, + 0.04162503033876419 + ], + "credit_map": [ + 0.020589428022503853, + 0.022534338757395744, + 0.02563389018177986, + 0.028802474960684776, + 0.020025277510285378, + 0.014925478026270866, + 0.03360030800104141, + 0.030727526172995567, + 0.005528594367206097, + 0.003648294135928154, + 0.009583480656147003, + 0.0077075473964214325, + 0.0049523417837917805, + 0.004851662088185549, + 0.012031705118715763, + 0.011396226473152637, + 0.0024186712689697742, + 0.0024491134099662304, + 0.003441235050559044, + 0.0026988512836396694, + 0.002566059585660696, + 0.0018072938546538353, + 0.004024326801300049, + 0.0027664825320243835, + 0.003684170078486204, + 0.0026435961481183767, + 0.004772222135215998, + 0.0021864157170057297, + 0.002635907381772995, + 0.007557292468845844, + 0.0016389836091548204, + 0.0034156846813857555, + 0.0037717940285801888, + 0.007342732045799494, + 0.006237619090825319, + 0.004707395564764738, + 0.006507093086838722, + 0.005156844388693571, + 0.011949973180890083, + 0.008329326286911964, + 0.002169347833842039, + 0.002835153602063656, + 0.008033379912376404, + 0.00312400353141129, + 0.0038067384157329798, + 0.002965520368888974, + 0.011882832273840904, + 0.003154539270326495, + 0.006301273126155138, + 0.003925642929971218, + 0.0058222427032887936, + 0.004292682278901339, + 0.004325577523559332, + 0.0041789161041378975, + 0.00376457953825593, + 0.0074151926673948765, + 0.03223617374897003, + 0.08749495446681976, + 0.07773590087890625, + 0.020339693874120712, + 0.1534254252910614, + 0.03435804694890976, + 0.021996019408106804, + 0.02097577415406704, + 0.005878274794667959, + 0.009493663907051086, + 0.006634114775806665, + 0.0161568783223629, + 0.009001933969557285, + 0.004805153701454401, + 0.01320688147097826, + 0.005390870850533247, + 0.006605012807995081, + 0.0042374590411782265, + 0.004164030309766531, + 0.005931050982326269, + 0.0053307069465518, + 0.007022263947874308, + 0.004807377699762583, + 0.003527135821059346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.02638811059296131 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003624657168984413 + ], + "opposite_logprob_tokens": [ + -5.628624439239502 + ], + "logprob_reward": -0.003624657168984413, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003624657168984413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0616137832403183, + 0.02059486135840416, + 0.020823946222662926, + 0.16409286856651306, + 0.049089979380369186, + 0.0783768817782402, + 0.0931568294763565, + 0.41550034284591675, + 0.026790158823132515, + 0.06996027380228043 + ], + "credit_map": [ + 0.009450208395719528, + 0.005990844685584307, + 0.00920194573700428, + 0.008230365812778473, + 0.006121362093836069, + 0.005216652061790228, + 0.009289318695664406, + 0.008113089017570019, + 0.0019736902322620153, + 0.001638818415813148, + 0.003427894553169608, + 0.002787749981507659, + 0.0019773144740611315, + 0.0018285838887095451, + 0.002535739680752158, + 0.004425070248544216, + 0.002796758431941271, + 0.0025900863111019135, + 0.002982790581882, + 0.001475697266869247, + 0.0026616447139531374, + 0.002158361952751875, + 0.0022063925862312317, + 0.003952214494347572, + 0.02788139507174492, + 0.005706450901925564, + 0.03455168381333351, + 0.008408721536397934, + 0.007387623190879822, + 0.06326986104249954, + 0.003118677530437708, + 0.013768465258181095, + 0.0038916992489248514, + 0.005296110641211271, + 0.007962256669998169, + 0.005221208557486534, + 0.0059731085784733295, + 0.0050656902603805065, + 0.008859892375767231, + 0.006820015609264374, + 0.003817915916442871, + 0.005649769678711891, + 0.023295696824789047, + 0.004621200263500214, + 0.0057107494212687016, + 0.002704824088141322, + 0.022067060694098473, + 0.010509672574698925, + 0.009888417087495327, + 0.01376873254776001, + 0.01376342959702015, + 0.011839209124445915, + 0.01328844390809536, + 0.010758713819086552, + 0.010996718890964985, + 0.00885317474603653, + 0.02951805852353573, + 0.07855931669473648, + 0.07814361900091171, + 0.019540414214134216, + 0.1383335292339325, + 0.03119824454188347, + 0.019205106422305107, + 0.021002070978283882, + 0.0022594660986214876, + 0.0030833298806101084, + 0.0022731239441782236, + 0.006546539720147848, + 0.0034651693422347307, + 0.0017476158682256937, + 0.005483692046254873, + 0.0019312223885208368, + 0.0107982587069273, + 0.008432535454630852, + 0.005415209103375673, + 0.014497315511107445, + 0.010514071211218834, + 0.008026708848774433, + 0.008103200234472752, + 0.004172982648015022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.003624657168984413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011798333143815398 + ], + "opposite_logprob_tokens": [ + -6.7511796951293945 + ], + "logprob_reward": -0.0011798333143815398, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011798333143815398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04320710524916649, + 0.04803323745727539, + 0.04731111600995064, + 0.26738932728767395, + 0.04748554527759552, + 0.10874205827713013, + 0.09100919216871262, + 0.06486827880144119, + 0.021440433338284492, + 0.26051366329193115 + ], + "credit_map": [ + 0.005006093531847, + 0.003566585248336196, + 0.00857487227767706, + 0.005429015029221773, + 0.0035806940868496895, + 0.0036030933260917664, + 0.006234827451407909, + 0.007211925927549601, + 0.005896804854273796, + 0.003553489688783884, + 0.006688012275844812, + 0.006520117167383432, + 0.0042997607961297035, + 0.0044358293525874615, + 0.008489033207297325, + 0.008150189183652401, + 0.005094816908240318, + 0.004597558174282312, + 0.006831997074186802, + 0.003907272592186928, + 0.0052828481420874596, + 0.004322173539549112, + 0.006978814024478197, + 0.01029563695192337, + 0.04659757763147354, + 0.00879501923918724, + 0.05412951111793518, + 0.013725112192332745, + 0.011511728167533875, + 0.10383186489343643, + 0.0068373121321201324, + 0.021961193531751633, + 0.0038768656086176634, + 0.006794421933591366, + 0.0071235541254282, + 0.004469886887818575, + 0.005969900172203779, + 0.005774341989308596, + 0.006442027632147074, + 0.007034544367343187, + 0.008205687627196312, + 0.012003548443317413, + 0.025829387828707695, + 0.006014078389853239, + 0.010822061449289322, + 0.006682600826025009, + 0.025792930275201797, + 0.013391765765845776, + 0.011459254659712315, + 0.01403643749654293, + 0.014422817155718803, + 0.011439193040132523, + 0.013917573727667332, + 0.007972940802574158, + 0.008282274007797241, + 0.009478699415922165, + 0.004676640033721924, + 0.008350971154868603, + 0.016219869256019592, + 0.004618547856807709, + 0.013076389208436012, + 0.0047311424277722836, + 0.004299638792872429, + 0.0088950851932168, + 0.0014587735058739781, + 0.0023621453437954187, + 0.002176858950406313, + 0.00620909221470356, + 0.0028230405878275633, + 0.001388302305713296, + 0.0034415984991937876, + 0.0015806197188794613, + 0.021785587072372437, + 0.042320944368839264, + 0.014457294717431068, + 0.0530928373336792, + 0.0437144935131073, + 0.04177488014101982, + 0.03096533566713333, + 0.012402274645864964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.0011798333143815398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47429898381233215 + ], + "opposite_logprob_tokens": [ + -0.9742989540100098 + ], + "logprob_reward": -0.47429898381233215, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.47429898381233215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04046891629695892, + 0.06951651722192764, + 0.0330161415040493, + 0.16310866177082062, + 0.026888957247138023, + 0.09480485320091248, + 0.05931776762008667, + 0.41157153248786926, + 0.021442675963044167, + 0.07986406236886978 + ], + "credit_map": [ + 0.004001036752015352, + 0.003995915874838829, + 0.006050865165889263, + 0.006772283464670181, + 0.003538471646606922, + 0.0030926004983484745, + 0.006137380842119455, + 0.006880354136228561, + 0.008068131282925606, + 0.002933862619102001, + 0.01539426390081644, + 0.005321898963302374, + 0.0038050974253565073, + 0.0073293354362249374, + 0.008118568919599056, + 0.018545346334576607, + 0.0036777728237211704, + 0.0033941371366381645, + 0.004633409436792135, + 0.0034154204186052084, + 0.003188641043379903, + 0.002903252374380827, + 0.0054120272397994995, + 0.006391474511474371, + 0.027872877195477486, + 0.005446615628898144, + 0.03447762504220009, + 0.008193224668502808, + 0.007006634026765823, + 0.06325887888669968, + 0.0029182075522840023, + 0.013934568502008915, + 0.0028712984640151262, + 0.0029808201361447573, + 0.0060372124426066875, + 0.0024862210266292095, + 0.003129109973087907, + 0.0022521335631608963, + 0.0028787923511117697, + 0.004253364633768797, + 0.004850981757044792, + 0.007169146090745926, + 0.0237420666962862, + 0.004692085552960634, + 0.00819623563438654, + 0.005208126734942198, + 0.030901845544576645, + 0.01004435308277607, + 0.0058716158382594585, + 0.009231144562363625, + 0.017904410138726234, + 0.003928081132471561, + 0.006505309604108334, + 0.003833525814116001, + 0.00424438389018178, + 0.007799288723617792, + 0.030110087245702744, + 0.07879794389009476, + 0.07465013116598129, + 0.017721613869071007, + 0.1392020881175995, + 0.031044403091073036, + 0.01997995562851429, + 0.020065229386091232, + 0.0020686634816229343, + 0.002763159107416868, + 0.0021374605130404234, + 0.004881665110588074, + 0.002733084140345454, + 0.0014461430255323648, + 0.0034521359484642744, + 0.0019603592809289694, + 0.0054718912579119205, + 0.015434100292623043, + 0.004212287720292807, + 0.016327276825904846, + 0.012134360149502754, + 0.011709388345479965, + 0.01015445962548256, + 0.004420277662575245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.47429898381233215 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot from a slightly high angle captures a captivating, sharp time-lapse of a giant Victoria Amazon water lily in full bloom. The scene begins with a spiny, dark green bud with a bronze hue emerging from the still, dark water, its surface reflecting the ever-changing ambient light. In an accelerated yet seamless sequence, the tightly curled bud begins its transformation, which seems to unfold over the course of an entire day and night. The outer petals deliberately peel away, revealing the massive, pure white inner petals. These waxy, boat-shaped petals gracefully unfold layer by layer in a rhythmic, hypnotic dance, their pristine surfaces capturing the transitional light, from the cool blue of dawn, through the bright clarity of midday, to the warm golden hues of sunset. This movement culminates when the flower fully opens, revealing its intricate golden center—a dense corolla that creates a stunning, vibrant contrast with the stark white of the blooming flower, creating a serene and awe-inspiring spectacle of natural beauty.", + "seed": 20148, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_2/index_104_seed_20148_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.106869736535009, + "vlm_logprob": -0.106869736535009, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.106869736535009, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.053972940891981125 + ], + "opposite_logprob_tokens": [ + -3.1789729595184326 + ], + "logprob_reward": -0.053972940891981125, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.053972940891981125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1985996514558792, + 0.0971633717417717, + 0.07468992471694946, + 0.024635741487145424, + 0.08592723309993744, + 0.10357953608036041, + 0.0549139603972435, + 0.14069093763828278, + 0.17962756752967834, + 0.04017205536365509 + ], + "credit_map": [ + 0.03112199157476425, + 0.014269950799643993, + 0.03195231407880783, + 0.012005270458757877, + 0.04749089479446411, + 0.01905982755124569, + 0.02504134736955166, + 0.017658066004514694, + 0.01926417462527752, + 0.00885598361492157, + 0.023092888295650482, + 0.006631145253777504, + 0.007509525865316391, + 0.009652797132730484, + 0.011986338533461094, + 0.010170514695346355, + 0.008742566220462322, + 0.010778959840536118, + 0.011698099784553051, + 0.005198897328227758, + 0.006695402320474386, + 0.011409706436097622, + 0.007895291782915592, + 0.012271000072360039, + 0.002547320444136858, + 0.0036843796260654926, + 0.004033942241221666, + 0.0021033037919551134, + 0.00407490273937583, + 0.002898646052926779, + 0.002636156976222992, + 0.0026570914778858423, + 0.01856919378042221, + 0.007204675581306219, + 0.009413481689989567, + 0.0055884406901896, + 0.021429380401968956, + 0.007474234793335199, + 0.007710677571594715, + 0.008537150919437408, + 0.013630551286041737, + 0.007898976095020771, + 0.014760144054889679, + 0.008600758388638496, + 0.01319160871207714, + 0.014751908369362354, + 0.009297835640609264, + 0.02144775167107582, + 0.009069857187569141, + 0.007149962708353996, + 0.005876834969967604, + 0.004339745733886957, + 0.01453743502497673, + 0.004217718727886677, + 0.004956630989909172, + 0.004765776917338371, + 0.03280472010374069, + 0.030036775395274162, + 0.01016900036484003, + 0.008999649435281754, + 0.013162452727556229, + 0.009976004250347614, + 0.009081815369427204, + 0.02646050974726677, + 0.01596858724951744, + 0.05416274443268776, + 0.005627007223665714, + 0.009524306282401085, + 0.07374028116464615, + 0.007415472995489836, + 0.008723273873329163, + 0.004465905483812094, + 0.006069566123187542, + 0.006631967145949602, + 0.0029922141693532467, + 0.003759619779884815, + 0.011710983701050282, + 0.0030716315377503633, + 0.0030160043388605118, + 0.002920068334788084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.053972940891981125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003210154827684164 + ], + "opposite_logprob_tokens": [ + -5.753210067749023 + ], + "logprob_reward": -0.003210154827684164, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003210154827684164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3200885057449341, + 0.03214698284864426, + 0.04311201348900795, + 0.047086864709854126, + 0.11244469881057739, + 0.07009516656398773, + 0.1372864693403244, + 0.0771075114607811, + 0.11811909079551697, + 0.04251272976398468 + ], + "credit_map": [ + 0.0635034590959549, + 0.01809118315577507, + 0.05524356663227081, + 0.019351577386260033, + 0.06650038063526154, + 0.028143854811787605, + 0.03564427047967911, + 0.03361018747091293, + 0.0041268630884587765, + 0.0038502663373947144, + 0.005006673280149698, + 0.0025945042725652456, + 0.004027333576232195, + 0.004639844875782728, + 0.0042946492321789265, + 0.0036068481858819723, + 0.004817247856408358, + 0.0047583500854671, + 0.00739890756085515, + 0.002571397228166461, + 0.004706304520368576, + 0.008205854333937168, + 0.0036148156505078077, + 0.007039136253297329, + 0.008347123861312866, + 0.004582569468766451, + 0.00623357342556119, + 0.0032078747171908617, + 0.004738735966384411, + 0.00721424026414752, + 0.003945157863199711, + 0.008817585185170174, + 0.01655467599630356, + 0.01422516442835331, + 0.010758429765701294, + 0.005368185229599476, + 0.03936381638050079, + 0.009482056833803654, + 0.00977958645671606, + 0.006912791170179844, + 0.006605738308280706, + 0.007251618895679712, + 0.008935333229601383, + 0.0070393760688602924, + 0.0072807022370398045, + 0.009958788752555847, + 0.012700237333774567, + 0.01032337173819542, + 0.023237893357872963, + 0.021507564932107925, + 0.008869016543030739, + 0.0049993302673101425, + 0.05376114696264267, + 0.008423076011240482, + 0.008452479727566242, + 0.008035967126488686, + 0.017013438045978546, + 0.015906617045402527, + 0.0047056605108082294, + 0.0059498511254787445, + 0.006870934274047613, + 0.0072866929695010185, + 0.006368360482156277, + 0.013005958870053291, + 0.01309886947274208, + 0.032284677028656006, + 0.003983876667916775, + 0.006626518908888102, + 0.04587443172931671, + 0.00504003930836916, + 0.007063275668770075, + 0.004147403873503208, + 0.004123946651816368, + 0.00976460613310337, + 0.0019947041291743517, + 0.0035217287950217724, + 0.014589517377316952, + 0.002555655315518379, + 0.0038796798326075077, + 0.0020828901324421167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.003210154827684164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019456282025203109 + ], + "opposite_logprob_tokens": [ + -6.251945495605469 + ], + "logprob_reward": -0.0019456282025203109, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019456282025203109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15814825892448425, + 0.07083127647638321, + 0.03882089629769325, + 0.05652647092938423, + 0.05959201604127884, + 0.2139439731836319, + 0.13033293187618256, + 0.09972455352544785, + 0.08829061686992645, + 0.08378899097442627 + ], + "credit_map": [ + 0.02216370403766632, + 0.0103351641446352, + 0.038352835923433304, + 0.006544194649904966, + 0.03498769551515579, + 0.013476819731295109, + 0.016679301857948303, + 0.015608548186719418, + 0.014380970038473606, + 0.0064451321959495544, + 0.022415053099393845, + 0.0033556916750967503, + 0.003437630133703351, + 0.00949324294924736, + 0.006111945025622845, + 0.005191608797758818, + 0.0036561042070388794, + 0.006757352501153946, + 0.0071115437895059586, + 0.004180254880338907, + 0.002996929455548525, + 0.004470390267670155, + 0.004699951969087124, + 0.004948366899043322, + 0.012259295210242271, + 0.005626915488392115, + 0.007107115816324949, + 0.003730805590748787, + 0.005984751973301172, + 0.005776519421488047, + 0.005497462581843138, + 0.01054360345005989, + 0.00958515889942646, + 0.0052038501016795635, + 0.007255925796926022, + 0.00431590061634779, + 0.015611523762345314, + 0.006813776213675737, + 0.005151152145117521, + 0.005654729902744293, + 0.025142230093479156, + 0.019814331084489822, + 0.0339057557284832, + 0.010576843284070492, + 0.027484335005283356, + 0.021084358915686607, + 0.016030482947826385, + 0.05990563705563545, + 0.02418421022593975, + 0.018797406926751137, + 0.009040082804858685, + 0.004608094226568937, + 0.048684343695640564, + 0.008394324220716953, + 0.00787395890802145, + 0.008750522509217262, + 0.024548066779971123, + 0.017914293333888054, + 0.008468102663755417, + 0.007349954918026924, + 0.00887687224894762, + 0.008255147375166416, + 0.008173847571015358, + 0.01613827608525753, + 0.006961904000490904, + 0.02060733363032341, + 0.003903144272044301, + 0.008325072936713696, + 0.02956763282418251, + 0.00444834167137742, + 0.011218756437301636, + 0.0032584310974925756, + 0.013247021473944187, + 0.016090722754597664, + 0.006354998331516981, + 0.009413021616637707, + 0.02031729370355606, + 0.004497698042541742, + 0.00964170228689909, + 0.004226528108119965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0019456282025203109 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009201106731779873 + ], + "opposite_logprob_tokens": [ + -7.000920295715332 + ], + "logprob_reward": -0.0009201106731779873, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009201106731779873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.155122771859169, + 0.033557720482349396, + 0.024570362642407417, + 0.03479847311973572, + 0.07896055281162262, + 0.10166354477405548, + 0.19925309717655182, + 0.09037921577692032, + 0.0677611455321312, + 0.21393315494060516 + ], + "credit_map": [ + 0.023364625871181488, + 0.009383936412632465, + 0.02993759512901306, + 0.007750191260129213, + 0.03025376796722412, + 0.02189665660262108, + 0.018920287489891052, + 0.013615699484944344, + 0.007158102001994848, + 0.0033788129221647978, + 0.00995231606066227, + 0.002178529743105173, + 0.0020044927950948477, + 0.003345789620652795, + 0.0033683544024825096, + 0.002171317348256707, + 0.002083609579131007, + 0.0026422967202961445, + 0.004675055854022503, + 0.0020300925243645906, + 0.0022178960498422384, + 0.0031455508433282375, + 0.0028528249822556973, + 0.004923033993691206, + 0.006782637443393469, + 0.003368810983374715, + 0.004405583720654249, + 0.0022013820707798004, + 0.004285142291337252, + 0.0031508691608905792, + 0.003733862191438675, + 0.006870180834084749, + 0.01115820650011301, + 0.01164389681071043, + 0.006910898722708225, + 0.0027676247991621494, + 0.03109820745885372, + 0.004791634622961283, + 0.006295444909483194, + 0.004294633865356445, + 0.012436727993190289, + 0.008797764778137207, + 0.015945833176374435, + 0.006261767819523811, + 0.011925854720175266, + 0.010703681036829948, + 0.00802649836987257, + 0.027565406635403633, + 0.03197990730404854, + 0.029012924060225487, + 0.016476338729262352, + 0.011532234959304333, + 0.0699901133775711, + 0.011887254193425179, + 0.012611746788024902, + 0.015762565657496452, + 0.011246129870414734, + 0.011046330444514751, + 0.010294935666024685, + 0.008276157081127167, + 0.007801724597811699, + 0.010768203064799309, + 0.011902905069291592, + 0.01904282346367836, + 0.004513112362474203, + 0.016079531982541084, + 0.0032193567603826523, + 0.008736214600503445, + 0.02006341516971588, + 0.003172453725710511, + 0.009393841959536076, + 0.002583215245977044, + 0.02947739139199257, + 0.04209015145897865, + 0.00960593018680811, + 0.023143064230680466, + 0.059925757348537445, + 0.01397280115634203, + 0.020006632432341576, + 0.01571141742169857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0009201106731779873 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742998480796814 + ], + "opposite_logprob_tokens": [ + -0.9742998480796814 + ], + "logprob_reward": -0.4742998480796814, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742998480796814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06085147708654404, + 0.1409822702407837, + 0.041229650378227234, + 0.021460825577378273, + 0.19259971380233765, + 0.12157368659973145, + 0.021402351558208466, + 0.2288486361503601, + 0.09231946617364883, + 0.07873186469078064 + ], + "credit_map": [ + 0.01269715465605259, + 0.005673975218087435, + 0.007349989376962185, + 0.00470315758138895, + 0.008667553775012493, + 0.005941582378000021, + 0.008458996191620827, + 0.0073590693064033985, + 0.03161183372139931, + 0.012357501313090324, + 0.04897039011120796, + 0.006519046146422625, + 0.006791342049837112, + 0.013116834685206413, + 0.013220683671534061, + 0.008394647389650345, + 0.005936083849519491, + 0.006410459987819195, + 0.006896322593092918, + 0.0032169048208743334, + 0.003946786746382713, + 0.0050368462689220905, + 0.0046478575095534325, + 0.0051383888348937035, + 0.0027385088615119457, + 0.0034055865835398436, + 0.0025915883015841246, + 0.0015283169923350215, + 0.0027766521088778973, + 0.0027323728427290916, + 0.0022184017580002546, + 0.0034693991765379906, + 0.04314430430531502, + 0.01882486790418625, + 0.018335996195673943, + 0.00929988268762827, + 0.06490462273359299, + 0.012378882616758347, + 0.013149470090866089, + 0.012561704032123089, + 0.015067057684063911, + 0.009585165418684483, + 0.01788160763680935, + 0.01119766291230917, + 0.012829530984163284, + 0.01368617545813322, + 0.01422891952097416, + 0.027097558602690697, + 0.0022597110364586115, + 0.0035191199276596308, + 0.0034471291583031416, + 0.0027799480594694614, + 0.0032413278240710497, + 0.001431057695299387, + 0.001923984382301569, + 0.0028000723104923964, + 0.04271474480628967, + 0.05436604097485542, + 0.014938710257411003, + 0.019115742295980453, + 0.02791704423725605, + 0.020012911409139633, + 0.022114481776952744, + 0.027668951079249382, + 0.006582970265299082, + 0.02468068338930607, + 0.005561379715800285, + 0.007070906460285187, + 0.03170609474182129, + 0.006142789963632822, + 0.006270783953368664, + 0.004303859546780586, + 0.010558054782450199, + 0.014136450365185738, + 0.00803174078464508, + 0.012456988915801048, + 0.012952462770044804, + 0.004847850184887648, + 0.009919501841068268, + 0.00582881411537528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.4742998480796814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot from a slightly high angle captures a captivating, sharp time-lapse of a giant Victoria Amazon water lily in full bloom. The scene begins with a spiny, dark green bud with a bronze hue emerging from the still, dark water, its surface reflecting the ever-changing ambient light. In an accelerated yet seamless sequence, the tightly curled bud begins its transformation, which seems to unfold over the course of an entire day and night. The outer petals deliberately peel away, revealing the massive, pure white inner petals. These waxy, boat-shaped petals gracefully unfold layer by layer in a rhythmic, hypnotic dance, their pristine surfaces capturing the transitional light, from the cool blue of dawn, through the bright clarity of midday, to the warm golden hues of sunset. This movement culminates when the flower fully opens, revealing its intricate golden center—a dense corolla that creates a stunning, vibrant contrast with the stark white of the blooming flower, creating a serene and awe-inspiring spectacle of natural beauty.", + "seed": 20149, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_2/index_104_seed_20149_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.18788271995726974, + "vlm_logprob": -0.18788271995726974, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.18788271995726974, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03225262835621834 + ], + "opposite_logprob_tokens": [ + -3.657252550125122 + ], + "logprob_reward": -0.03225262835621834, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03225262835621834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22956161201000214, + 0.06722825765609741, + 0.03936414793133736, + 0.041126858443021774, + 0.05002661049365997, + 0.022710571065545082, + 0.031389910727739334, + 0.4355749487876892, + 0.03148593381047249, + 0.051531098783016205 + ], + "credit_map": [ + 0.04415988549590111, + 0.014248890802264214, + 0.014921378344297409, + 0.03139062598347664, + 0.02118268795311451, + 0.019979126751422882, + 0.05492763593792915, + 0.028751397505402565, + 0.010631526820361614, + 0.006522101350128651, + 0.008207479491829872, + 0.007321702782064676, + 0.0065487767569720745, + 0.007396353874355555, + 0.00891682505607605, + 0.01168349850922823, + 0.0035594089422374964, + 0.004547550342977047, + 0.004428751766681671, + 0.004955982323735952, + 0.0043519060127437115, + 0.004305364564061165, + 0.007101582828909159, + 0.006113602314144373, + 0.004521800670772791, + 0.00407581590116024, + 0.006990151014178991, + 0.004001457244157791, + 0.004990908317267895, + 0.006786736659705639, + 0.004728363361209631, + 0.005031629931181669, + 0.004271899349987507, + 0.0032775315921753645, + 0.009795531630516052, + 0.004223242402076721, + 0.003980525769293308, + 0.008524922654032707, + 0.005148847587406635, + 0.010804113000631332, + 0.0028533935546875, + 0.002565626287832856, + 0.0030425339937210083, + 0.00217945734038949, + 0.002641732571646571, + 0.003113111015409231, + 0.002203434705734253, + 0.0041112820617854595, + 0.004425693769007921, + 0.003016932401806116, + 0.0037989916745573282, + 0.0037026002537459135, + 0.002947852248325944, + 0.004255376290529966, + 0.0036361573729664087, + 0.0056063104420900345, + 0.07923831790685654, + 0.07293061912059784, + 0.018774360418319702, + 0.025282539427280426, + 0.17277240753173828, + 0.025676051154732704, + 0.023442592471837997, + 0.017458057031035423, + 0.003060393501073122, + 0.00344489305280149, + 0.0034416653215885162, + 0.0027799387462437153, + 0.002408720087260008, + 0.005623877514153719, + 0.003339965594932437, + 0.007386485114693642, + 0.009958751499652863, + 0.0054922145791351795, + 0.003508752677589655, + 0.0040890914388000965, + 0.006106480024755001, + 0.005017284769564867, + 0.004627991933375597, + 0.012730536982417107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.03225262835621834 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005965997464954853 + ], + "opposite_logprob_tokens": [ + -5.1309661865234375 + ], + "logprob_reward": -0.005965997464954853, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005965997464954853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24075070023536682, + 0.21561722457408905, + 0.09982232749462128, + 0.026545900851488113, + 0.10478459298610687, + 0.032217204570770264, + 0.05707338824868202, + 0.056969910860061646, + 0.07846904546022415, + 0.08774964511394501 + ], + "credit_map": [ + 0.042631205171346664, + 0.01443877536803484, + 0.014563203789293766, + 0.041841261088848114, + 0.01721215806901455, + 0.021073894575238228, + 0.05844317004084587, + 0.03054703213274479, + 0.040997959673404694, + 0.012042508460581303, + 0.013419833034276962, + 0.019944941624999046, + 0.015478012152016163, + 0.01604263111948967, + 0.06268670409917831, + 0.0350046306848526, + 0.012687033973634243, + 0.016684845089912415, + 0.007861320860683918, + 0.015247457660734653, + 0.008609156124293804, + 0.009104644879698753, + 0.016805509105324745, + 0.012822361662983894, + 0.0021527709905058146, + 0.003333963453769684, + 0.0036347638815641403, + 0.004871368873864412, + 0.0033424971625208855, + 0.00347212515771389, + 0.002584554022178054, + 0.003153855912387371, + 0.009462096728384495, + 0.007396678440272808, + 0.01723744533956051, + 0.007340461015701294, + 0.010478015057742596, + 0.02423728071153164, + 0.008625658228993416, + 0.02000696212053299, + 0.003965470008552074, + 0.003568819724023342, + 0.0036449602339416742, + 0.0027703314553946257, + 0.0038394734729081392, + 0.004335567820817232, + 0.0031461247708648443, + 0.006946458015590906, + 0.006565084680914879, + 0.006838835310190916, + 0.00403487728908658, + 0.006148134358227253, + 0.011026425287127495, + 0.01030744332820177, + 0.004809433128684759, + 0.007343158591538668, + 0.009385890327394009, + 0.008100447244942188, + 0.004392486531287432, + 0.00372215430252254, + 0.010421118699014187, + 0.0047501446679234505, + 0.003260699100792408, + 0.0129369692876935, + 0.009333830326795578, + 0.008096644654870033, + 0.007768062874674797, + 0.005294305272400379, + 0.008308535441756248, + 0.02334463596343994, + 0.005712875165045261, + 0.010610162280499935, + 0.014504079706966877, + 0.00597296841442585, + 0.008087962865829468, + 0.005747121758759022, + 0.007241393905133009, + 0.021011926233768463, + 0.004624740686267614, + 0.020559456199407578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005965997464954853 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013403251068666577 + ], + "opposite_logprob_tokens": [ + -6.626340389251709 + ], + "logprob_reward": -0.0013403251068666577, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013403251068666577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17447106540203094, + 0.07657996565103531, + 0.06138252466917038, + 0.03291005641222, + 0.13085712492465973, + 0.0654696673154831, + 0.09030703455209732, + 0.21599745750427246, + 0.027523554861545563, + 0.12450157105922699 + ], + "credit_map": [ + 0.018438756465911865, + 0.014149611815810204, + 0.011320983991026878, + 0.033232104033231735, + 0.026836877688765526, + 0.031036891043186188, + 0.02049902454018593, + 0.018956808373332024, + 0.009059710428118706, + 0.006443735212087631, + 0.007512506563216448, + 0.007909982465207577, + 0.007258099503815174, + 0.017176596447825432, + 0.012033119797706604, + 0.009186217561364174, + 0.0071687716990709305, + 0.010571413673460484, + 0.004198450595140457, + 0.00923690665513277, + 0.006169030908495188, + 0.00520444056019187, + 0.010335863567888737, + 0.008497646078467369, + 0.003546170424669981, + 0.004827367141842842, + 0.005594389513134956, + 0.003556052688509226, + 0.0029405902605503798, + 0.004589262884110212, + 0.004431634675711393, + 0.0034245906863361597, + 0.013411535881459713, + 0.011633098125457764, + 0.019200997427105904, + 0.010928231291472912, + 0.01212947815656662, + 0.031099367886781693, + 0.014369305223226547, + 0.018085112795233727, + 0.007450885139405727, + 0.009228612296283245, + 0.007347438484430313, + 0.005558334290981293, + 0.007444400805979967, + 0.007651981431990862, + 0.008762724697589874, + 0.012025289237499237, + 0.01602729596197605, + 0.012192444875836372, + 0.01061982661485672, + 0.009261095896363258, + 0.009202348999679089, + 0.012783009558916092, + 0.007381923496723175, + 0.012839091010391712, + 0.012247048318386078, + 0.03842347115278244, + 0.00812953244894743, + 0.013993784785270691, + 0.08310912549495697, + 0.017288176342844963, + 0.00920543260872364, + 0.03360089287161827, + 0.0028665720019489527, + 0.005142510868608952, + 0.002902007894590497, + 0.0029643517918884754, + 0.003950011450797319, + 0.004348136018961668, + 0.0028250899631530046, + 0.002524877432733774, + 0.023250006139278412, + 0.010533805936574936, + 0.008394964970648289, + 0.008897172287106514, + 0.006471391301602125, + 0.027797099202871323, + 0.006366988178342581, + 0.03279014676809311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.0013403251068666577 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015119798481464386 + ], + "opposite_logprob_tokens": [ + -6.501512050628662 + ], + "logprob_reward": -0.0015119798481464386, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015119798481464386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12644745409488678, + 0.07518071681261063, + 0.07034006714820862, + 0.05825420841574669, + 0.0601586252450943, + 0.0277392715215683, + 0.03335525840520859, + 0.21645590662956238, + 0.07738654315471649, + 0.25468188524246216 + ], + "credit_map": [ + 0.027070006355643272, + 0.009559331461787224, + 0.007421784568578005, + 0.019434891641139984, + 0.007884851656854153, + 0.0074206567369401455, + 0.03415604308247566, + 0.013499891385436058, + 0.008603411726653576, + 0.005719041917473078, + 0.008918647654354572, + 0.009106486104428768, + 0.0059968335554003716, + 0.013925746083259583, + 0.009531041607260704, + 0.013379506766796112, + 0.008683268912136555, + 0.010616011917591095, + 0.005579968448728323, + 0.008713391609489918, + 0.00606095464900136, + 0.00621603662148118, + 0.01386251486837864, + 0.010607916861772537, + 0.006392753217369318, + 0.007178771309554577, + 0.008472412824630737, + 0.006469545885920525, + 0.007664720993489027, + 0.009397108107805252, + 0.006318204570561647, + 0.006360693834722042, + 0.006780529860407114, + 0.004177324939519167, + 0.013689425773918629, + 0.004606080707162619, + 0.005246800370514393, + 0.007477002218365669, + 0.005367701407521963, + 0.012813761830329895, + 0.0029767598025500774, + 0.002383521990850568, + 0.004361629486083984, + 0.0021553575061261654, + 0.002785453572869301, + 0.003966402728110552, + 0.003271005814895034, + 0.005839141551405191, + 0.004962787963449955, + 0.0049963220953941345, + 0.002830832963809371, + 0.003797908080741763, + 0.0037404168397188187, + 0.006334366742521524, + 0.0032622171565890312, + 0.0034304072614759207, + 0.046451348811388016, + 0.030917752534151077, + 0.010571151971817017, + 0.012255861423909664, + 0.07228337973356247, + 0.013325272127985954, + 0.011864018626511097, + 0.018787121400237083, + 0.007871460169553757, + 0.010991680435836315, + 0.010858394205570221, + 0.008297980763018131, + 0.007700649090111256, + 0.009964493103325367, + 0.0064917150884866714, + 0.015210178680717945, + 0.0237684715539217, + 0.037252336740493774, + 0.02930588461458683, + 0.029485180974006653, + 0.026185493916273117, + 0.05186188966035843, + 0.02212863601744175, + 0.034693993628025055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.0015119798481464386 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983426690101624 + ], + "opposite_logprob_tokens": [ + -0.5233426690101624 + ], + "logprob_reward": -0.8983426690101624, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983426690101624, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05648557096719742, + 0.0773608535528183, + 0.024449262768030167, + 0.040719591081142426, + 0.03741278126835823, + 0.02931799553334713, + 0.03178127482533455, + 0.4382754862308502, + 0.1085168719291687, + 0.1556803584098816 + ], + "credit_map": [ + 0.00760284811258316, + 0.00597321568056941, + 0.00701009389013052, + 0.011505907401442528, + 0.0035471576265990734, + 0.006890591699630022, + 0.006558234803378582, + 0.007397520821541548, + 0.01218785997480154, + 0.00407937029376626, + 0.005044699180871248, + 0.008404484018683434, + 0.005368156358599663, + 0.006255693268030882, + 0.022986402735114098, + 0.01303419005125761, + 0.002695626812055707, + 0.002682922873646021, + 0.0026224630419164896, + 0.003088164608925581, + 0.0022845349740236998, + 0.0027304990217089653, + 0.0049654957838356495, + 0.0033795563504099846, + 0.005542607046663761, + 0.0033459069672971964, + 0.005258840508759022, + 0.0027756968047469854, + 0.005600834731012583, + 0.007804068271070719, + 0.005008668638765812, + 0.005382969044148922, + 0.003433267818763852, + 0.004223845899105072, + 0.004423198755830526, + 0.004977116361260414, + 0.004162071272730827, + 0.003941804636269808, + 0.0064061847515404224, + 0.00584529060870409, + 0.0034332324285060167, + 0.0032612220384180546, + 0.003143435576930642, + 0.0027030387427657843, + 0.003369684563949704, + 0.004785981960594654, + 0.003463162574917078, + 0.005158238112926483, + 0.0026184648741036654, + 0.005965826567262411, + 0.0032601167913526297, + 0.00424857996404171, + 0.004025880713015795, + 0.004585689399391413, + 0.003957748878747225, + 0.0031189678702503443, + 0.0780499204993248, + 0.07384348660707474, + 0.01892612688243389, + 0.025368180125951767, + 0.17487336695194244, + 0.025775622576475143, + 0.023598896339535713, + 0.017839891836047173, + 0.008720751851797104, + 0.01958487555384636, + 0.008659856393933296, + 0.00951570738106966, + 0.013704361394047737, + 0.027765125036239624, + 0.009746218100190163, + 0.010819974355399609, + 0.024297073483467102, + 0.01578333228826523, + 0.016895407810807228, + 0.011373120360076427, + 0.013909958302974701, + 0.033506982028484344, + 0.012646929360926151, + 0.027267547324299812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.8983426690101624 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static medium shot from a slightly high angle captures a captivating, sharp time-lapse of a giant Victoria Amazon water lily in full bloom. The scene begins with a spiny, dark green bud with a bronze hue emerging from the still, dark water, its surface reflecting the ever-changing ambient light. In an accelerated yet seamless sequence, the tightly curled bud begins its transformation, which seems to unfold over the course of an entire day and night. The outer petals deliberately peel away, revealing the massive, pure white inner petals. These waxy, boat-shaped petals gracefully unfold layer by layer in a rhythmic, hypnotic dance, their pristine surfaces capturing the transitional light, from the cool blue of dawn, through the bright clarity of midday, to the warm golden hues of sunset. This movement culminates when the flower fully opens, revealing its intricate golden center—a dense corolla that creates a stunning, vibrant contrast with the stark white of the blooming flower, creating a serene and awe-inspiring spectacle of natural beauty.", + "seed": 20150, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_2/index_104_seed_20150_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.189539594668895, + "vlm_logprob": -0.189539594668895, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.189539594668895, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08737719058990479 + ], + "opposite_logprob_tokens": [ + -2.5873770713806152 + ], + "logprob_reward": -0.08737719058990479, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08737719058990479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18707552552223206, + 0.06962990760803223, + 0.051947206258773804, + 0.08286450803279877, + 0.09409406036138535, + 0.08384045213460922, + 0.033797524869441986, + 0.2624621093273163, + 0.06822625547647476, + 0.06606248766183853 + ], + "credit_map": [ + 0.026923302561044693, + 0.015609105117619038, + 0.040319763123989105, + 0.023259488865733147, + 0.016516661271452904, + 0.026204440742731094, + 0.020297138020396233, + 0.017945634201169014, + 0.010414359159767628, + 0.00561376940459013, + 0.011259987950325012, + 0.006119232624769211, + 0.006006142590194941, + 0.00875481404364109, + 0.008900726214051247, + 0.012560874223709106, + 0.007381584029644728, + 0.006819911766797304, + 0.006565926130861044, + 0.008291750214993954, + 0.004578813444823027, + 0.006067659240216017, + 0.006198592018336058, + 0.006042971275746822, + 0.01241249218583107, + 0.015075130388140678, + 0.010355459526181221, + 0.010081345215439796, + 0.010474002920091152, + 0.009232604876160622, + 0.007827355526387691, + 0.007406117860227823, + 0.014444563537836075, + 0.008008659817278385, + 0.014560975134372711, + 0.010278507135808468, + 0.00894058309495449, + 0.014078151434659958, + 0.011475111357867718, + 0.012307503260672092, + 0.016219697892665863, + 0.007918855175375938, + 0.017196033149957657, + 0.006890907883644104, + 0.007680654525756836, + 0.011690867133438587, + 0.005289989989250898, + 0.01095344964414835, + 0.004021044354885817, + 0.003926921170204878, + 0.003596852533519268, + 0.004030116833746433, + 0.0041888258419930935, + 0.005632633809000254, + 0.004630591254681349, + 0.003770538605749607, + 0.035785406827926636, + 0.007455299608409405, + 0.011332806199789047, + 0.010325184091925621, + 0.0146251255646348, + 0.090607650578022, + 0.018239371478557587, + 0.07409127056598663, + 0.008298378437757492, + 0.0030172152910381556, + 0.014744111336767673, + 0.0032296853605657816, + 0.0028067314997315407, + 0.00711015472188592, + 0.0036295230966061354, + 0.02539045736193657, + 0.00969164539128542, + 0.005102687049657106, + 0.0034567788243293762, + 0.006748679559677839, + 0.008988610468804836, + 0.01037419680505991, + 0.00967591255903244, + 0.01202397421002388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.08737719058990479 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023326972499489784 + ], + "opposite_logprob_tokens": [ + -3.773326873779297 + ], + "logprob_reward": -0.023326972499489784, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023326972499489784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23667727410793304, + 0.07826384902000427, + 0.09778422862291336, + 0.09589481353759766, + 0.05940205976366997, + 0.05605204030871391, + 0.06150586158037186, + 0.0425059013068676, + 0.17515428364276886, + 0.0967596247792244 + ], + "credit_map": [ + 0.029298443347215652, + 0.02698272280395031, + 0.04644074663519859, + 0.032474637031555176, + 0.027240151539444923, + 0.027441123500466347, + 0.02773277275264263, + 0.019066670909523964, + 0.01124491635710001, + 0.009754735045135021, + 0.01282087154686451, + 0.008368594571948051, + 0.005509008187800646, + 0.009900983422994614, + 0.008527026511728764, + 0.012137709185481071, + 0.016120653599500656, + 0.00918751023709774, + 0.019549638032913208, + 0.009428254328668118, + 0.006547207944095135, + 0.01036035642027855, + 0.010683894157409668, + 0.015906712040305138, + 0.019097434356808662, + 0.009686578065156937, + 0.012140667997300625, + 0.01095095369964838, + 0.00847080908715725, + 0.013146040961146355, + 0.007402015384286642, + 0.015000313520431519, + 0.00878352764993906, + 0.0067717814818024635, + 0.00770124327391386, + 0.008283503353595734, + 0.006155847106128931, + 0.008049925789237022, + 0.0058759115636348724, + 0.007780318148434162, + 0.009638911113142967, + 0.006153488531708717, + 0.008310114033520222, + 0.005510180257260799, + 0.0075548854656517506, + 0.006157275289297104, + 0.004245162010192871, + 0.008482026867568493, + 0.006856692023575306, + 0.007331008091568947, + 0.005460123065859079, + 0.007381186820566654, + 0.006887495983392, + 0.011700350791215897, + 0.009914053604006767, + 0.005974951665848494, + 0.007869568653404713, + 0.0032498876098543406, + 0.00288893049582839, + 0.0032680893782526255, + 0.005788119044154882, + 0.006659786216914654, + 0.006053078453987837, + 0.006728442385792732, + 0.025732629001140594, + 0.00983358919620514, + 0.01895947940647602, + 0.008319624699652195, + 0.01041257381439209, + 0.02959466353058815, + 0.011476625688374043, + 0.06082512065768242, + 0.017397990450263023, + 0.0056433831341564655, + 0.00608085049316287, + 0.008711077272891998, + 0.006699463352560997, + 0.019812149927020073, + 0.010068175382912159, + 0.022346528246998787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.023326972499489784 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076359291560947895 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.0076359291560947895, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076359291560947895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10540974885225296, + 0.06371080130338669, + 0.051737453788518906, + 0.1318121701478958, + 0.10001332312822342, + 0.13630039989948273, + 0.09796831756830215, + 0.0418829508125782, + 0.08413873612880707, + 0.18702606856822968 + ], + "credit_map": [ + 0.011388610117137432, + 0.013765182346105576, + 0.018508290871977806, + 0.016293292865157127, + 0.0107039799913764, + 0.013229708187282085, + 0.0111978305503726, + 0.010322866030037403, + 0.007379347458481789, + 0.009556475095450878, + 0.008842449635267258, + 0.007636184338480234, + 0.0069948649033904076, + 0.008090421557426453, + 0.00780426524579525, + 0.007406797260046005, + 0.007939480245113373, + 0.006541721057146788, + 0.008094125427305698, + 0.009191717952489853, + 0.005309281405061483, + 0.0042764730751514435, + 0.006465772166848183, + 0.0039188917726278305, + 0.02169489674270153, + 0.0188458189368248, + 0.019821463152766228, + 0.017963090911507607, + 0.015273166820406914, + 0.014521848410367966, + 0.013475784100592136, + 0.01021612249314785, + 0.014945227652788162, + 0.010442416183650494, + 0.01577153429389, + 0.013979635201394558, + 0.009609142318367958, + 0.012538541108369827, + 0.0104805463925004, + 0.01224629208445549, + 0.025361845269799232, + 0.010684504173696041, + 0.02522183768451214, + 0.011219742707908154, + 0.00988584291189909, + 0.02050888165831566, + 0.00912874098867178, + 0.02428903616964817, + 0.011792179197072983, + 0.010365955531597137, + 0.012707357294857502, + 0.01124553382396698, + 0.011000476777553558, + 0.01496875286102295, + 0.013118278235197067, + 0.01276978850364685, + 0.00601614685729146, + 0.0020477978978306055, + 0.0024013647343963385, + 0.002582377754151821, + 0.0027897555846720934, + 0.007602636702358723, + 0.006400787737220526, + 0.012042088434100151, + 0.008841192349791527, + 0.0036469041369855404, + 0.01847894676029682, + 0.0042547136545181274, + 0.003577620256692171, + 0.010523633100092411, + 0.00466770539060235, + 0.03014802746474743, + 0.03490360081195831, + 0.010483840480446815, + 0.011618335731327534, + 0.017871223390102386, + 0.020623931661248207, + 0.04487083852291107, + 0.02473735623061657, + 0.021916963160037994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.0076359291560947895 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031943507492542267 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.0031943507492542267, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031943507492542267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08109118044376373, + 0.04679858684539795, + 0.03659311681985855, + 0.08762398362159729, + 0.17930887639522552, + 0.07777680456638336, + 0.05855775997042656, + 0.0891035869717598, + 0.08180007338523865, + 0.26134607195854187 + ], + "credit_map": [ + 0.011982380412518978, + 0.009034006856381893, + 0.01742401532828808, + 0.012089884839951992, + 0.007383429445326328, + 0.006232793442904949, + 0.009186120703816414, + 0.007758548483252525, + 0.008000953122973442, + 0.0044348714873194695, + 0.007998060435056686, + 0.0036497300025075674, + 0.0036216939333826303, + 0.005128301214426756, + 0.005449392832815647, + 0.008515586145222187, + 0.005148848984390497, + 0.006073513068258762, + 0.0048246830701828, + 0.006523570511490107, + 0.003009395906701684, + 0.003396552987396717, + 0.004475472029298544, + 0.0031410795636475086, + 0.010631839744746685, + 0.01574990339577198, + 0.009934499859809875, + 0.01080307736992836, + 0.010319666005671024, + 0.009702116250991821, + 0.009615755639970303, + 0.01086712907999754, + 0.030386067926883698, + 0.014761670492589474, + 0.03306789696216583, + 0.02370329387485981, + 0.015082674100995064, + 0.023295406252145767, + 0.018164027482271194, + 0.020847829058766365, + 0.014931907877326012, + 0.0053990851156413555, + 0.015960833057761192, + 0.0063597108237445354, + 0.005172400269657373, + 0.013045004568994045, + 0.0054060472175478935, + 0.011501818895339966, + 0.009255788289010525, + 0.0052202134393155575, + 0.007856590673327446, + 0.005394252482801676, + 0.009868943132460117, + 0.008156446740031242, + 0.005879970733076334, + 0.006925554480403662, + 0.011620127595961094, + 0.003080192720517516, + 0.00395483523607254, + 0.003884070785716176, + 0.004571516998112202, + 0.019534753635525703, + 0.012970157898962498, + 0.02948792837560177, + 0.010861534625291824, + 0.0040365164168179035, + 0.019928675144910812, + 0.004034733399748802, + 0.0051316008903086185, + 0.013108490034937859, + 0.005690786987543106, + 0.01900773122906685, + 0.03879086673259735, + 0.02567419782280922, + 0.017317038029432297, + 0.023344095796346664, + 0.04064028337597847, + 0.05227521434426308, + 0.03382158279418945, + 0.029482783749699593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.0031943507492542267 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video show a spiny, dark green bud with a bronze hue emerging from still, dark water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the outer petals peel away to reveal massive, pure white inner petals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the flower fully open to reveal an intricate golden center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lighting transition from cool blue dawn to bright midday and then to warm golden sunset?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261635303497314 + ], + "opposite_logprob_tokens": [ + -0.5761635303497314 + ], + "logprob_reward": -0.8261635303497314, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261635303497314, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11036570370197296, + 0.0858471468091011, + 0.07591487467288971, + 0.10752110928297043, + 0.18366488814353943, + 0.09305094182491302, + 0.05485440790653229, + 0.07255486398935318, + 0.1076696589589119, + 0.10855646431446075 + ], + "credit_map": [ + 0.015442385338246822, + 0.009086227044463158, + 0.025656092911958694, + 0.010515205562114716, + 0.010923027992248535, + 0.012114581651985645, + 0.012412196025252342, + 0.014215986244380474, + 0.012338941916823387, + 0.008514821529388428, + 0.011536323465406895, + 0.0056145223788917065, + 0.009488889016211033, + 0.015565820038318634, + 0.008736980147659779, + 0.014050846919417381, + 0.009156300686299801, + 0.013895111158490181, + 0.008061977103352547, + 0.009154275991022587, + 0.009157329797744751, + 0.007839890196919441, + 0.009796874597668648, + 0.00885311234742403, + 0.019609572365880013, + 0.015194887295365334, + 0.014593543484807014, + 0.012526694685220718, + 0.010517156682908535, + 0.012292519211769104, + 0.01019043754786253, + 0.01259628590196371, + 0.036876313388347626, + 0.014953874051570892, + 0.028362205252051353, + 0.01858007349073887, + 0.01617417298257351, + 0.02693425863981247, + 0.017843903973698616, + 0.02394009195268154, + 0.014613400213420391, + 0.009468292817473412, + 0.016973823308944702, + 0.00997124146670103, + 0.009140883572399616, + 0.013169513083994389, + 0.007381863426417112, + 0.012331928126513958, + 0.006629568990319967, + 0.006156241055577993, + 0.007810192182660103, + 0.0055330414324998856, + 0.005053075961768627, + 0.009331412613391876, + 0.007210954558104277, + 0.007129922974854708, + 0.008727207779884338, + 0.002519670408219099, + 0.004158989991992712, + 0.0033406659495085478, + 0.004811673425137997, + 0.01983771100640297, + 0.007537790108472109, + 0.021621156483888626, + 0.01107376255095005, + 0.004127779509872198, + 0.02477295882999897, + 0.005364588461816311, + 0.0046601914800703526, + 0.013778249733150005, + 0.005683432798832655, + 0.03820869326591492, + 0.023611079901456833, + 0.005409356206655502, + 0.008167778141796589, + 0.007470679935067892, + 0.009696073830127716, + 0.02851860783994198, + 0.01008264534175396, + 0.01560024544596672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.8261635303497314 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_3/index_tensor([413, 413])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_3/index_tensor([413, 413])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4e6fa96cd27c422a848cec412d11eb7b848fe9ff --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_3/index_tensor([413, 413])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static wide-angle shot captures a full-body view of a clump of iron-arrow bamboo, rendered entirely in the stark, high-contrast black-and-white ink style of Charles Burns. The camera remains completely still, rendering the scene like a horrifying panel from a graphic novel. Amidst the pure, inky void, the bamboo vines don't just wither; they twist into unnatural, harrowing shapes, their surfaces riddled with grotesque, tumorous protrusions and deep, shadowy pits that defy their natural form. Leaves curl inward like the dehydrated, arthritic fingers of a corpse, each a sickly, flat sheet of white. True to the designated style, each leaf vein is meticulously rendered as a sharp black line against the sickly white, emphasizing the fragile, lifeless texture. The harsh lighting, with no discernible source, creates a world of absolute binary opposition—no grays, no soft gradients, only solid black shapes and stark, unforgiving white surfaces. This high-contrast aesthetic creates a deeply unsettling atmosphere of bio-horror and decay, transforming the plants into giant, static specimens of botanical mutation.", + "seed": 25250, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_3/index_413_seed_25250_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05254433564841747, + "vlm_logprob": -0.05254433564841747, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05254433564841747, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10549015551805496 + ], + "opposite_logprob_tokens": [ + -2.355490207672119 + ], + "logprob_reward": -0.10549015551805496, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10549015551805496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46602684259414673, + 0.10373517125844955, + 0.05055156722664833, + 0.0500849224627018, + 0.05473192036151886, + 0.057386770844459534, + 0.04066275805234909, + 0.049560677260160446, + 0.030331846326589584, + 0.0969274565577507 + ], + "credit_map": [ + 0.048930492252111435, + 0.031439509242773056, + 0.045294392853975296, + 0.07319815456867218, + 0.04219185560941696, + 0.041309699416160583, + 0.13910458981990814, + 0.04455816000699997, + 0.011304658837616444, + 0.008563465438783169, + 0.007315563503652811, + 0.02067830041050911, + 0.018652508035302162, + 0.011530517600476742, + 0.014863303862512112, + 0.01082684751600027, + 0.007245913147926331, + 0.004479146096855402, + 0.003949443809688091, + 0.006908793468028307, + 0.007323199883103371, + 0.006035233847796917, + 0.00790358241647482, + 0.006706260144710541, + 0.00563844433054328, + 0.003668781602755189, + 0.0036198513116687536, + 0.007354752626270056, + 0.010169667191803455, + 0.007424273993819952, + 0.006968486588448286, + 0.0052406638860702515, + 0.008937900885939598, + 0.0038312566466629505, + 0.004988717380911112, + 0.0067533706314861774, + 0.00822812132537365, + 0.008175954222679138, + 0.007798115722835064, + 0.006018484942615032, + 0.007784522138535976, + 0.004613400902599096, + 0.0039610834792256355, + 0.009212900884449482, + 0.008927770890295506, + 0.009516056627035141, + 0.007514151744544506, + 0.0058568865060806274, + 0.0066382125951349735, + 0.0037675066851079464, + 0.0049801976419985294, + 0.00617040041834116, + 0.005180359818041325, + 0.004686109721660614, + 0.005446984898298979, + 0.0037929867394268513, + 0.005756983533501625, + 0.004523886367678642, + 0.00397869385778904, + 0.008256034925580025, + 0.009327389299869537, + 0.006819521076977253, + 0.004961859434843063, + 0.005936307366937399, + 0.004175725392997265, + 0.0023845029063522816, + 0.003110770136117935, + 0.004641624633222818, + 0.004654445685446262, + 0.004212320316582918, + 0.00395806273445487, + 0.0031943931244313717, + 0.008526304736733437, + 0.007143413182348013, + 0.00899302028119564, + 0.034172508865594864, + 0.012389757670462132, + 0.006612675730139017, + 0.007523030042648315, + 0.01156674511730671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.10549015551805496 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610168255865574 + ], + "opposite_logprob_tokens": [ + -4.141101837158203 + ], + "logprob_reward": -0.01610168255865574, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610168255865574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24756446480751038, + 0.1282215267419815, + 0.152125746011734, + 0.06740887463092804, + 0.08807754516601562, + 0.04433261230587959, + 0.019931955263018608, + 0.08608489483594894, + 0.03776136040687561, + 0.12849095463752747 + ], + "credit_map": [ + 0.019298527389764786, + 0.012431379407644272, + 0.010204051621258259, + 0.08257840573787689, + 0.0762389600276947, + 0.011006874032318592, + 0.022913914173841476, + 0.01289236918091774, + 0.024232037365436554, + 0.010614192113280296, + 0.010818252339959145, + 0.01742163673043251, + 0.014263897202908993, + 0.01386212557554245, + 0.01306213904172182, + 0.023947255685925484, + 0.02198234386742115, + 0.010438555851578712, + 0.013044300489127636, + 0.029651127755641937, + 0.016653528437018394, + 0.01982175186276436, + 0.020002366974949837, + 0.0205317921936512, + 0.007565981242805719, + 0.006021247711032629, + 0.004502727184444666, + 0.008280632086098194, + 0.012665528804063797, + 0.010558762587606907, + 0.009753257036209106, + 0.008060747757554054, + 0.011552602052688599, + 0.008303530514240265, + 0.008451742120087147, + 0.016945281997323036, + 0.00979909859597683, + 0.012563248164951801, + 0.009502771310508251, + 0.01095927320420742, + 0.006254578474909067, + 0.0037036158610135317, + 0.0035940867383033037, + 0.01041453704237938, + 0.005440721288323402, + 0.005706511903554201, + 0.00429344130679965, + 0.004925125744193792, + 0.0030484506860375404, + 0.0018465315224602818, + 0.002177569782361388, + 0.0037039436865597963, + 0.0026003201492130756, + 0.002606843365356326, + 0.0019906351808458567, + 0.0019576610065996647, + 0.013417156413197517, + 0.008141551166772842, + 0.0065474980510771275, + 0.012711500748991966, + 0.01628093048930168, + 0.010845587588846684, + 0.00952543318271637, + 0.008615242317318916, + 0.006735759787261486, + 0.003783028107136488, + 0.003320568474009633, + 0.005230036098510027, + 0.006152368150651455, + 0.004712455905973911, + 0.0036714731249958277, + 0.00415567634627223, + 0.01173313520848751, + 0.01024080254137516, + 0.011585948057472706, + 0.04149008169770241, + 0.01864955574274063, + 0.011172604747116566, + 0.012828209437429905, + 0.010790619999170303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.01610168255865574 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076340362429618835 + ], + "opposite_logprob_tokens": [ + -4.882634162902832 + ], + "logprob_reward": -0.0076340362429618835, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076340362429618835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25655534863471985, + 0.16427446901798248, + 0.07460817694664001, + 0.12078302353620529, + 0.039466388523578644, + 0.07875125110149384, + 0.04487619549036026, + 0.04308243840932846, + 0.03734590858221054, + 0.14025680720806122 + ], + "credit_map": [ + 0.02956469915807247, + 0.018598197028040886, + 0.016589047387242317, + 0.05152149498462677, + 0.06207209452986717, + 0.017073338851332664, + 0.024519773200154305, + 0.03661670535802841, + 0.018216896802186966, + 0.011188681237399578, + 0.015530460514128208, + 0.03609449043869972, + 0.024063602089881897, + 0.018810054287314415, + 0.015353185124695301, + 0.025017106905579567, + 0.010083924047648907, + 0.005359667353332043, + 0.00663677928969264, + 0.011481333523988724, + 0.011477110907435417, + 0.010301641188561916, + 0.010000837966799736, + 0.009266875684261322, + 0.010046216659247875, + 0.008268876932561398, + 0.006201904732733965, + 0.0160988699644804, + 0.026740195229649544, + 0.02271646074950695, + 0.01833421364426613, + 0.012376291677355766, + 0.004197927191853523, + 0.003119995119050145, + 0.0034769203048199415, + 0.005623544100672007, + 0.006314451340585947, + 0.006622758693993092, + 0.004496077541261911, + 0.005614711437374353, + 0.00961621105670929, + 0.006417116150259972, + 0.005850363988429308, + 0.016410037875175476, + 0.011977041140198708, + 0.011685915291309357, + 0.007548402529209852, + 0.00924616027623415, + 0.0054387920536100864, + 0.004453917033970356, + 0.004550587385892868, + 0.0057366993278265, + 0.007042297162115574, + 0.007249796763062477, + 0.0054938350804150105, + 0.004910273011773825, + 0.005009748507291079, + 0.004109791945666075, + 0.0033991243690252304, + 0.007315266411751509, + 0.0038946897257119417, + 0.008251272141933441, + 0.004260914400219917, + 0.006841631140559912, + 0.005272646900266409, + 0.004466806072741747, + 0.003026561578735709, + 0.00401379493996501, + 0.005317743867635727, + 0.006769252009689808, + 0.004713011439889669, + 0.0037660913076251745, + 0.01142093539237976, + 0.010398448444902897, + 0.011332354508340359, + 0.03878755867481232, + 0.028885865584015846, + 0.011593353003263474, + 0.016485024243593216, + 0.011353266425430775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0076340362429618835 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020582307130098343 + ], + "opposite_logprob_tokens": [ + -3.8955821990966797 + ], + "logprob_reward": -0.020582307130098343, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020582307130098343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2836056351661682, + 0.10747653245925903, + 0.09076008200645447, + 0.07765305787324905, + 0.08670088648796082, + 0.12977752089500427, + 0.029262907803058624, + 0.047569647431373596, + 0.06524372100830078, + 0.08194995671510696 + ], + "credit_map": [ + 0.02223423682153225, + 0.015577391721308231, + 0.019076509401202202, + 0.08363484591245651, + 0.07691720128059387, + 0.021014941856265068, + 0.025241078808903694, + 0.019909456372261047, + 0.016380760818719864, + 0.011446702294051647, + 0.008966753259301186, + 0.02382192574441433, + 0.013502094894647598, + 0.010773017071187496, + 0.01111580803990364, + 0.01146948803216219, + 0.01379786990582943, + 0.006738933268934488, + 0.007469178177416325, + 0.013101098127663136, + 0.009983833879232407, + 0.011583790183067322, + 0.010936830192804337, + 0.017148561775684357, + 0.0077004460617899895, + 0.005269824061542749, + 0.00529657443985343, + 0.007982955314218998, + 0.016739757731556892, + 0.013957168906927109, + 0.012390433810651302, + 0.008315900340676308, + 0.009298407472670078, + 0.005344792269170284, + 0.007270232308655977, + 0.01688353344798088, + 0.013077656738460064, + 0.013464540243148804, + 0.009804321452975273, + 0.011557415127754211, + 0.015174140222370625, + 0.008804747834801674, + 0.010576453991234303, + 0.0265842005610466, + 0.02198607847094536, + 0.016536815091967583, + 0.014065547846257687, + 0.01604955829679966, + 0.002771198283880949, + 0.0029415104072541, + 0.0022606842685490847, + 0.00412878580391407, + 0.004410234279930592, + 0.004754404537379742, + 0.0031659682281315327, + 0.004830125253647566, + 0.006633775308728218, + 0.003731783013790846, + 0.005186695605516434, + 0.00727486377581954, + 0.007597237825393677, + 0.006284736096858978, + 0.005395667627453804, + 0.005464895162731409, + 0.007546567358076572, + 0.007757949642837048, + 0.005524250213056803, + 0.00971142016351223, + 0.007618846837431192, + 0.012332857586443424, + 0.006142378319054842, + 0.008609459735453129, + 0.007915227673947811, + 0.00877342652529478, + 0.007540487218648195, + 0.01412164606153965, + 0.01792135462164879, + 0.008266018703579903, + 0.00999821163713932, + 0.007413596846163273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.020582307130098343 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291349679231644 + ], + "opposite_logprob_tokens": [ + -2.2379136085510254 + ], + "logprob_reward": -0.11291349679231644, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291349679231644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38132989406585693, + 0.05745717138051987, + 0.04107794165611267, + 0.1321340650320053, + 0.029856737703084946, + 0.030760997906327248, + 0.05882187560200691, + 0.09792936593294144, + 0.04569774121046066, + 0.12493427842855453 + ], + "credit_map": [ + 0.03332100808620453, + 0.024923447519540787, + 0.03042139858007431, + 0.07319972664117813, + 0.08959674835205078, + 0.030515778809785843, + 0.06485960632562637, + 0.034492138773202896, + 0.007793067488819361, + 0.004721958190202713, + 0.005044438876211643, + 0.010644990019500256, + 0.00543418200686574, + 0.008606897667050362, + 0.008142835460603237, + 0.007068785838782787, + 0.005125630181282759, + 0.003680758411064744, + 0.003990985918790102, + 0.005555358715355396, + 0.006242882460355759, + 0.005681905895471573, + 0.005012062378227711, + 0.005788350943475962, + 0.012447793036699295, + 0.009112064726650715, + 0.008147032931447029, + 0.018785787746310234, + 0.02543877437710762, + 0.022936567664146423, + 0.0189238078892231, + 0.01634221151471138, + 0.004026011098176241, + 0.002795300679281354, + 0.0024935787077993155, + 0.00507636833935976, + 0.0037579357158392668, + 0.00440744124352932, + 0.0035136800725013018, + 0.0037864167243242264, + 0.004125479143112898, + 0.0017759583424776793, + 0.0023278496228158474, + 0.004858715459704399, + 0.004174104426056147, + 0.0050211516208946705, + 0.0042052543722093105, + 0.004272480495274067, + 0.007012588903307915, + 0.0046377116814255714, + 0.005879714619368315, + 0.011502808891236782, + 0.009746991097927094, + 0.007160206791013479, + 0.006643348373472691, + 0.006238499190658331, + 0.012088155373930931, + 0.008231229148805141, + 0.00804068148136139, + 0.017642684280872345, + 0.015346341766417027, + 0.014258142560720444, + 0.010653086937963963, + 0.011669027619063854, + 0.006258415523916483, + 0.005227233748883009, + 0.004032299388200045, + 0.009293830022215843, + 0.005969668738543987, + 0.005581538658589125, + 0.004570492077618837, + 0.004764257930219173, + 0.009582414291799068, + 0.008840261958539486, + 0.011823094449937344, + 0.048537544906139374, + 0.01977197825908661, + 0.0065505895763635635, + 0.009434855543076992, + 0.010393519885838032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.11291349679231644 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide-angle shot captures a full-body view of a clump of iron-arrow bamboo, rendered entirely in the stark, high-contrast black-and-white ink style of Charles Burns. The camera remains completely still, rendering the scene like a horrifying panel from a graphic novel. Amidst the pure, inky void, the bamboo vines don't just wither; they twist into unnatural, harrowing shapes, their surfaces riddled with grotesque, tumorous protrusions and deep, shadowy pits that defy their natural form. Leaves curl inward like the dehydrated, arthritic fingers of a corpse, each a sickly, flat sheet of white. True to the designated style, each leaf vein is meticulously rendered as a sharp black line against the sickly white, emphasizing the fragile, lifeless texture. The harsh lighting, with no discernible source, creates a world of absolute binary opposition—no grays, no soft gradients, only solid black shapes and stark, unforgiving white surfaces. This high-contrast aesthetic creates a deeply unsettling atmosphere of bio-horror and decay, transforming the plants into giant, static specimens of botanical mutation.", + "seed": 25251, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_3/index_413_seed_25251_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0885837584733963, + "vlm_logprob": -0.0885837584733963, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.0885837584733963, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04561490938067436 + ], + "opposite_logprob_tokens": [ + -3.1706149578094482 + ], + "logprob_reward": -0.04561490938067436, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04561490938067436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06885025650262833, + 0.3878040611743927, + 0.198075070977211, + 0.12528067827224731, + 0.01565811038017273, + 0.016505973413586617, + 0.04178817197680473, + 0.08934023976325989, + 0.02716967463493347, + 0.02952776663005352 + ], + "credit_map": [ + 0.00812750868499279, + 0.0074045658111572266, + 0.00982113927602768, + 0.011850254610180855, + 0.005262057762593031, + 0.010592789389193058, + 0.005638098809868097, + 0.010153849609196186, + 0.021228928118944168, + 0.016435690224170685, + 0.13658544421195984, + 0.15065671503543854, + 0.012090777978301048, + 0.014094190672039986, + 0.015525925904512405, + 0.02118641696870327, + 0.026760336011648178, + 0.01742199808359146, + 0.01712842285633087, + 0.08074446022510529, + 0.008804671466350555, + 0.016231754794716835, + 0.017552398145198822, + 0.013431034982204437, + 0.014859233051538467, + 0.010154563002288342, + 0.011694862507283688, + 0.015441619791090488, + 0.012068180367350578, + 0.02620781399309635, + 0.01599614880979061, + 0.01885826699435711, + 0.0019587588030844927, + 0.0016262950375676155, + 0.001867573824711144, + 0.0018667553085833788, + 0.0018373746424913406, + 0.002535108244046569, + 0.0016975350445136428, + 0.0022687118034809828, + 0.0016229144530370831, + 0.0027029921766370535, + 0.0015605352818965912, + 0.001831777743063867, + 0.001967550488188863, + 0.0027394630014896393, + 0.002190926345065236, + 0.0018898170674219728, + 0.0030421381816267967, + 0.0046933963894844055, + 0.00520183565095067, + 0.006570814643055201, + 0.004525312222540379, + 0.005405301693826914, + 0.0044660065323114395, + 0.007883370853960514, + 0.010096271522343159, + 0.005396777298301458, + 0.007459975313395262, + 0.031751569360494614, + 0.007523242849856615, + 0.004925839602947235, + 0.007869181223213673, + 0.014317394234240055, + 0.0034885494969785213, + 0.006177367176860571, + 0.0026849056594073772, + 0.002866670722141862, + 0.002276879968121648, + 0.003230966394767165, + 0.0026752883568406105, + 0.003769049420952797, + 0.004914354532957077, + 0.003876730799674988, + 0.0038211243227124214, + 0.00402330607175827, + 0.004197621252387762, + 0.0033134748227894306, + 0.002253487706184387, + 0.003127670381218195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.04561490938067436 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038139645010232925 + ], + "opposite_logprob_tokens": [ + -3.288139581680298 + ], + "logprob_reward": -0.038139645010232925, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038139645010232925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20229916274547577, + 0.17735399305820465, + 0.050803832709789276, + 0.09425479918718338, + 0.05913974717259407, + 0.019584618508815765, + 0.02754290960729122, + 0.1661594659090042, + 0.11985993385314941, + 0.08300157636404037 + ], + "credit_map": [ + 0.028813600540161133, + 0.026241159066557884, + 0.025225715711712837, + 0.020429974421858788, + 0.01376609317958355, + 0.027577344328165054, + 0.015077413991093636, + 0.04516785219311714, + 0.009851791895925999, + 0.007152191363275051, + 0.06400294601917267, + 0.07046603411436081, + 0.0045274109579622746, + 0.0059981569647789, + 0.005929653998464346, + 0.009425811469554901, + 0.011361820623278618, + 0.004381590522825718, + 0.005261213518679142, + 0.00584144564345479, + 0.0032795711886137724, + 0.0059579345397651196, + 0.005284192971885204, + 0.009436063468456268, + 0.013020619750022888, + 0.0062358626164495945, + 0.009094002656638622, + 0.012174558825790882, + 0.009591056033968925, + 0.015893761068582535, + 0.012050913646817207, + 0.016194025054574013, + 0.016779907047748566, + 0.006004868075251579, + 0.006155162584036589, + 0.004535730462521315, + 0.005294762551784515, + 0.008445176295936108, + 0.005479747895151377, + 0.006444394588470459, + 0.004322818946093321, + 0.0020296566653996706, + 0.0023265115451067686, + 0.002141473116353154, + 0.0025467940140515566, + 0.0021720887161791325, + 0.0016340294387191534, + 0.002411246532574296, + 0.003338415641337633, + 0.002664948580786586, + 0.0033061716239899397, + 0.003685483243316412, + 0.0034467775840312243, + 0.0035090786404907703, + 0.003675600280985236, + 0.003916432615369558, + 0.048733294010162354, + 0.013419180177152157, + 0.013606654480099678, + 0.02019651234149933, + 0.01935342140495777, + 0.01410173624753952, + 0.012613723985850811, + 0.02413495071232319, + 0.01827174797654152, + 0.04317723587155342, + 0.007710456382483244, + 0.00794314593076706, + 0.006067087408155203, + 0.00818826723843813, + 0.004959635902196169, + 0.023542357608675957, + 0.014265635050833225, + 0.009427985176444054, + 0.008565250784158707, + 0.013356329873204231, + 0.011285778135061264, + 0.009736569598317146, + 0.007705836556851864, + 0.00865818839520216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.038139645010232925 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633563131093979 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633563131093979, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633563131093979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1297575682401657, + 0.17081603407859802, + 0.09986404329538345, + 0.0724785327911377, + 0.13022051751613617, + 0.059573180973529816, + 0.1319350004196167, + 0.1299208551645279, + 0.045796461403369904, + 0.02963775396347046 + ], + "credit_map": [ + 0.014191817492246628, + 0.015761390328407288, + 0.016754666343331337, + 0.0189782977104187, + 0.010867638513445854, + 0.02110297977924347, + 0.010658446699380875, + 0.02144235000014305, + 0.011123653501272202, + 0.017530757933855057, + 0.045894745737314224, + 0.04894215613603592, + 0.008645557798445225, + 0.013110052794218063, + 0.011288783513009548, + 0.014280339702963829, + 0.008674928918480873, + 0.009801040403544903, + 0.011899271048605442, + 0.033377960324287415, + 0.005439484026283026, + 0.014108621515333652, + 0.008111108094453812, + 0.008451636880636215, + 0.008188705891370773, + 0.006924559362232685, + 0.006899862550199032, + 0.009701109491288662, + 0.0059989760629832745, + 0.011147101409733295, + 0.012446699663996696, + 0.01117152813822031, + 0.014414760284125805, + 0.017196644097566605, + 0.015070033259689808, + 0.016680395230650902, + 0.01441635750234127, + 0.02158172056078911, + 0.015257551334798336, + 0.015603079460561275, + 0.00880881305783987, + 0.009284330531954765, + 0.005390102509409189, + 0.006036562379449606, + 0.007092886604368687, + 0.008651681244373322, + 0.005744748283177614, + 0.008564063347876072, + 0.012681927531957626, + 0.01146742608398199, + 0.01569133624434471, + 0.011975935660302639, + 0.020392000675201416, + 0.022209353744983673, + 0.015208165161311626, + 0.022308876737952232, + 0.03506966307759285, + 0.008818630129098892, + 0.009464988484978676, + 0.010033243335783482, + 0.007514670491218567, + 0.01063208095729351, + 0.007216006517410278, + 0.041171588003635406, + 0.005894822534173727, + 0.00847841240465641, + 0.005102965515106916, + 0.005812954157590866, + 0.004999373108148575, + 0.006120150443166494, + 0.0046983216889202595, + 0.004689467139542103, + 0.0050447559915483, + 0.003947586752474308, + 0.003393807914108038, + 0.0037979751359671354, + 0.004171157255768776, + 0.003228458110243082, + 0.0030911238864064217, + 0.0029628919437527657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.007633563131093979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810659423470497 + ], + "opposite_logprob_tokens": [ + -3.288106679916382 + ], + "logprob_reward": -0.03810659423470497, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810659423470497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16927871108055115, + 0.11154190450906754, + 0.0785704255104065, + 0.08206497877836227, + 0.06164803355932236, + 0.05744556337594986, + 0.05814928561449051, + 0.16645467281341553, + 0.10243026912212372, + 0.11241614073514938 + ], + "credit_map": [ + 0.01931578479707241, + 0.023102127015590668, + 0.024633387103676796, + 0.025575384497642517, + 0.01459118165075779, + 0.02539316937327385, + 0.01493528950959444, + 0.021732378751039505, + 0.009871053509414196, + 0.011664977297186852, + 0.021227236837148666, + 0.026059823110699654, + 0.009186234325170517, + 0.009961127303540707, + 0.010717175900936127, + 0.01285427063703537, + 0.009552705101668835, + 0.006976904813200235, + 0.013300971128046513, + 0.01130647026002407, + 0.005710798781365156, + 0.008747859857976437, + 0.009278278797864914, + 0.013696434907615185, + 0.00889694131910801, + 0.007570948917418718, + 0.009392807260155678, + 0.01033722423017025, + 0.007556908763945103, + 0.01378809567540884, + 0.013082534074783325, + 0.011439517140388489, + 0.009316833689808846, + 0.006155411712825298, + 0.007158007472753525, + 0.007782922126352787, + 0.007280619349330664, + 0.007668325677514076, + 0.007488735485821962, + 0.008797180838882923, + 0.010494250804185867, + 0.00887705385684967, + 0.006102288141846657, + 0.005235705059021711, + 0.0064258393831551075, + 0.006511259824037552, + 0.004957776982337236, + 0.008841385133564472, + 0.0063853017054498196, + 0.005624893121421337, + 0.012070818804204464, + 0.006367628928273916, + 0.00799979642033577, + 0.007463223300874233, + 0.006212880834937096, + 0.00602474482730031, + 0.05676673352718353, + 0.010826357640326023, + 0.010827132500708103, + 0.03134419396519661, + 0.009845265187323093, + 0.010469838045537472, + 0.007924933917820454, + 0.028450213372707367, + 0.012904624454677105, + 0.024425053969025612, + 0.009792538359761238, + 0.00822059903293848, + 0.0074140941724181175, + 0.008032634854316711, + 0.008060411550104618, + 0.023580316454172134, + 0.01416630670428276, + 0.010801033116877079, + 0.01256466843187809, + 0.019297612830996513, + 0.016456399112939835, + 0.013125160709023476, + 0.01256166398525238, + 0.013443296775221825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.03810659423470497 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31342408061027527 + ], + "opposite_logprob_tokens": [ + -1.3134241104125977 + ], + "logprob_reward": -0.31342408061027527, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31342408061027527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07767961174249649, + 0.4138518273830414, + 0.09306392073631287, + 0.06333428621292114, + 0.043795738369226456, + 0.05009446665644646, + 0.0731964185833931, + 0.08238524943590164, + 0.057954803109169006, + 0.044643692672252655 + ], + "credit_map": [ + 0.009222385473549366, + 0.006762908771634102, + 0.008691915310919285, + 0.011369705200195312, + 0.005343778990209103, + 0.013878043740987778, + 0.005455511622130871, + 0.016955358907580376, + 0.023198630660772324, + 0.01769445091485977, + 0.1468963921070099, + 0.1619027554988861, + 0.012449502944946289, + 0.01398588065057993, + 0.015272241085767746, + 0.022451965138316154, + 0.015583780594170094, + 0.008310041390359402, + 0.010848489589989185, + 0.014723772183060646, + 0.0068045505322515965, + 0.011197423562407494, + 0.01106955949217081, + 0.014526305720210075, + 0.009334101341664791, + 0.006509785540401936, + 0.0068756467662751675, + 0.0075570507906377316, + 0.006078642327338457, + 0.009092655032873154, + 0.009949922561645508, + 0.007936484180390835, + 0.004851225763559341, + 0.005324255209416151, + 0.0051961238496005535, + 0.006336035672575235, + 0.004270640201866627, + 0.005700474139302969, + 0.0061218636110424995, + 0.005995121318846941, + 0.005483788438141346, + 0.005811106413602829, + 0.004947477951645851, + 0.006464127451181412, + 0.006889711134135723, + 0.008077533915638924, + 0.006653921213001013, + 0.005766797345131636, + 0.009445407427847385, + 0.011485476978123188, + 0.008568488992750645, + 0.009208070114254951, + 0.00738117890432477, + 0.007990104146301746, + 0.008143468759953976, + 0.010974223725497723, + 0.02307497337460518, + 0.006359780207276344, + 0.005867564585059881, + 0.00606786971911788, + 0.008234737440943718, + 0.006588631775230169, + 0.005618811585009098, + 0.020572884008288383, + 0.00606057234108448, + 0.013068378902971745, + 0.005714301485568285, + 0.005054880864918232, + 0.0060668340884149075, + 0.006550957448780537, + 0.006115506868809462, + 0.00932337436825037, + 0.00798623077571392, + 0.007629203610122204, + 0.006204860284924507, + 0.005276112351566553, + 0.00439497921615839, + 0.005078934133052826, + 0.0030412718188017607, + 0.005032100714743137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.31342408061027527 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide-angle shot captures a full-body view of a clump of iron-arrow bamboo, rendered entirely in the stark, high-contrast black-and-white ink style of Charles Burns. The camera remains completely still, rendering the scene like a horrifying panel from a graphic novel. Amidst the pure, inky void, the bamboo vines don't just wither; they twist into unnatural, harrowing shapes, their surfaces riddled with grotesque, tumorous protrusions and deep, shadowy pits that defy their natural form. Leaves curl inward like the dehydrated, arthritic fingers of a corpse, each a sickly, flat sheet of white. True to the designated style, each leaf vein is meticulously rendered as a sharp black line against the sickly white, emphasizing the fragile, lifeless texture. The harsh lighting, with no discernible source, creates a world of absolute binary opposition—no grays, no soft gradients, only solid black shapes and stark, unforgiving white surfaces. This high-contrast aesthetic creates a deeply unsettling atmosphere of bio-horror and decay, transforming the plants into giant, static specimens of botanical mutation.", + "seed": 25252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_3/index_413_seed_25252_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.20807889536954463, + "vlm_logprob": -0.20807889536954463, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.20807889536954463, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1829344481229782 + ], + "opposite_logprob_tokens": [ + -1.8079344034194946 + ], + "logprob_reward": -0.1829344481229782, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1829344481229782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5301613807678223, + 0.03704938665032387, + 0.027488429099321365, + 0.04070621728897095, + 0.024092016741633415, + 0.16560852527618408, + 0.035046763718128204, + 0.046394214034080505, + 0.06847667694091797, + 0.024976419284939766 + ], + "credit_map": [ + 0.05252774432301521, + 0.09051081538200378, + 0.036981649696826935, + 0.04354807734489441, + 0.14153477549552917, + 0.048034921288490295, + 0.042307302355766296, + 0.07471610605716705, + 0.003935197833925486, + 0.004647112917155027, + 0.006675418466329575, + 0.0037514835130423307, + 0.0065838005393743515, + 0.004078686702996492, + 0.0032380789052695036, + 0.004139604978263378, + 0.00276936125010252, + 0.003272801171988249, + 0.00613355590030551, + 0.0024870613124221563, + 0.0043535390868783, + 0.002828940749168396, + 0.0022861571051180363, + 0.0033570087980479, + 0.0058348895981907845, + 0.004759978968650103, + 0.007920634932816029, + 0.0042809690348804, + 0.00444208225235343, + 0.0045152874663472176, + 0.0038323651533573866, + 0.005120007786899805, + 0.0028056472074240446, + 0.0028647473081946373, + 0.0025923890061676502, + 0.003330114996060729, + 0.003933481406420469, + 0.0023427221458405256, + 0.002747597871348262, + 0.0034753161016851664, + 0.017117442563176155, + 0.012964431196451187, + 0.02797705866396427, + 0.028801705688238144, + 0.015417174436151981, + 0.013972107321023941, + 0.016136644408106804, + 0.03322196006774902, + 0.002985082333907485, + 0.004360757302492857, + 0.010817036963999271, + 0.003954469691962004, + 0.003534965217113495, + 0.0020432923920452595, + 0.0024136013817042112, + 0.0049375551752746105, + 0.005700760520994663, + 0.004999914206564426, + 0.007207465823739767, + 0.00701435562223196, + 0.005278713535517454, + 0.004992417059838772, + 0.005239191465079784, + 0.005961391609162092, + 0.016302205622196198, + 0.004622711334377527, + 0.02420017309486866, + 0.006568534299731255, + 0.0039006276056170464, + 0.0030800336971879005, + 0.005782299675047398, + 0.004020086955279112, + 0.003207507776096463, + 0.0022608921863138676, + 0.0036194585263729095, + 0.003459497354924679, + 0.0026193817611783743, + 0.003044815268367529, + 0.003328098217025399, + 0.003436767728999257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.1829344481229782 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328370571136475 + ], + "opposite_logprob_tokens": [ + -0.7578370571136475 + ], + "logprob_reward": -0.6328370571136475, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328370571136475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12746037542819977, + 0.19711193442344666, + 0.13183774054050446, + 0.0705753043293953, + 0.0699765756726265, + 0.11443722248077393, + 0.03187866508960724, + 0.06303896009922028, + 0.09973304718732834, + 0.09395024180412292 + ], + "credit_map": [ + 0.011891299858689308, + 0.025383897125720978, + 0.007573073264211416, + 0.009577597491443157, + 0.03576250374317169, + 0.010358672589063644, + 0.010631989687681198, + 0.016281340271234512, + 0.020446334034204483, + 0.023588359355926514, + 0.04064313694834709, + 0.020806826651096344, + 0.0316743329167366, + 0.021630503237247467, + 0.017102865502238274, + 0.021219559013843536, + 0.009548541158437729, + 0.015222011134028435, + 0.014705055393278599, + 0.01193571649491787, + 0.0229719877243042, + 0.01725820079445839, + 0.017834311351180077, + 0.022361911833286285, + 0.007172558456659317, + 0.0054918392561376095, + 0.014149216003715992, + 0.00685969227924943, + 0.006929050199687481, + 0.006407757289707661, + 0.005654947832226753, + 0.017910243943333626, + 0.006484163459390402, + 0.00803788285702467, + 0.013333679176867008, + 0.00793437659740448, + 0.015483411960303783, + 0.0062293391674757, + 0.00610763905569911, + 0.006366073619574308, + 0.01012637373059988, + 0.009341975674033165, + 0.022199194878339767, + 0.01917773298919201, + 0.012015202082693577, + 0.009460829198360443, + 0.009905456565320492, + 0.022210445255041122, + 0.0020506593864411116, + 0.006417022552341223, + 0.005516562145203352, + 0.003589833388105035, + 0.004822594579309225, + 0.0016725726891309023, + 0.0021705920808017254, + 0.00563882477581501, + 0.007956393994390965, + 0.006943094078451395, + 0.00832679308950901, + 0.007604317273944616, + 0.0062993247993290424, + 0.008952304720878601, + 0.0067062764428555965, + 0.0102504538372159, + 0.020768264308571815, + 0.007395661436021328, + 0.02918807603418827, + 0.011053783819079399, + 0.011590112932026386, + 0.005554484203457832, + 0.006243143230676651, + 0.00793951191008091, + 0.011678844690322876, + 0.010091787204146385, + 0.015462073497474194, + 0.01702631078660488, + 0.01055841613560915, + 0.007750371471047401, + 0.011253814212977886, + 0.010128619149327278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.6328370571136475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002491229446604848 + ], + "opposite_logprob_tokens": [ + -6.002490997314453 + ], + "logprob_reward": -0.002491229446604848, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002491229446604848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17193828523159027, + 0.09627753496170044, + 0.08325549960136414, + 0.07582470029592514, + 0.08776687830686569, + 0.10906685888767242, + 0.0633239820599556, + 0.060570843517780304, + 0.17826609313488007, + 0.07370935380458832 + ], + "credit_map": [ + 0.03269149735569954, + 0.02299206331372261, + 0.018383467569947243, + 0.013070688582956791, + 0.022820882499217987, + 0.013967565260827541, + 0.0129657918587327, + 0.03504631668329239, + 0.018733305856585503, + 0.0077380589209496975, + 0.014845984056591988, + 0.009045642800629139, + 0.009713757783174515, + 0.012991371564567089, + 0.008677110075950623, + 0.014532293193042278, + 0.007822448387742043, + 0.010517389513552189, + 0.010220179334282875, + 0.008458289317786694, + 0.01961769536137581, + 0.006407619453966618, + 0.011608932167291641, + 0.008602939546108246, + 0.007154336664825678, + 0.011365173384547234, + 0.008440492674708366, + 0.013170850463211536, + 0.01209014281630516, + 0.008535563014447689, + 0.007137014996260405, + 0.007931116968393326, + 0.008149486035108566, + 0.012700389139354229, + 0.0110065508633852, + 0.009817084297537804, + 0.018957549706101418, + 0.007986847311258316, + 0.00964618381112814, + 0.009502783417701721, + 0.01058296300470829, + 0.009405197575688362, + 0.01346158143132925, + 0.02072404883801937, + 0.012262609787285328, + 0.011246801353991032, + 0.01081218384206295, + 0.02057146653532982, + 0.004238965921103954, + 0.015159417875111103, + 0.006575013976544142, + 0.004080494865775108, + 0.004165970254689455, + 0.006109881680458784, + 0.004170140251517296, + 0.01882409304380417, + 0.007773516699671745, + 0.006343873683363199, + 0.010922975838184357, + 0.009298206306993961, + 0.006734580732882023, + 0.006176376715302467, + 0.006154991220682859, + 0.00716632092371583, + 0.04217381402850151, + 0.011975121684372425, + 0.06924235075712204, + 0.01610640622675419, + 0.010683692060410976, + 0.007056488189846277, + 0.011477201245725155, + 0.009551012888550758, + 0.009251996874809265, + 0.0077120037749409676, + 0.012311427854001522, + 0.011824837885797024, + 0.010170567780733109, + 0.006242028437554836, + 0.007302397862076759, + 0.008894085884094238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.002491229446604848 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586160942912102 + ], + "opposite_logprob_tokens": [ + -3.8955862522125244 + ], + "logprob_reward": -0.020586160942912102, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586160942912102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15978045761585236, + 0.08627332001924515, + 0.11535584181547165, + 0.08649708330631256, + 0.08619949966669083, + 0.09893227368593216, + 0.05266309902071953, + 0.08754723519086838, + 0.13148221373558044, + 0.09526902437210083 + ], + "credit_map": [ + 0.03223962336778641, + 0.018531423062086105, + 0.01896780915558338, + 0.012302474118769169, + 0.03758830577135086, + 0.013473829254508018, + 0.010060546919703484, + 0.016616452485322952, + 0.015969982370734215, + 0.009188410826027393, + 0.013475918211042881, + 0.010393357835710049, + 0.010511461645364761, + 0.007568224333226681, + 0.009728682227432728, + 0.009437281638383865, + 0.008484596386551857, + 0.010994983837008476, + 0.02207949012517929, + 0.011661338619887829, + 0.023908067494630814, + 0.01282588392496109, + 0.012137829326093197, + 0.013263641856610775, + 0.013151784427464008, + 0.01138339377939701, + 0.015129338018596172, + 0.010167781263589859, + 0.012282886542379856, + 0.008535471744835377, + 0.0061660036444664, + 0.009680419228971004, + 0.008128073066473007, + 0.013545775786042213, + 0.009667206555604935, + 0.01260079350322485, + 0.01596694253385067, + 0.007287857122719288, + 0.009354167617857456, + 0.009648682549595833, + 0.013444948010146618, + 0.00993905309587717, + 0.011535577476024628, + 0.01444809790700674, + 0.015056414529681206, + 0.00990857370197773, + 0.008881965652108192, + 0.01571764424443245, + 0.005120750050991774, + 0.008021717891097069, + 0.006493684370070696, + 0.00616465462371707, + 0.006441966164857149, + 0.004359490238130093, + 0.005072491243481636, + 0.010988344438374043, + 0.011093762703239918, + 0.00893300399184227, + 0.011869160458445549, + 0.009069511666893959, + 0.011208979412913322, + 0.010527869686484337, + 0.010109695605933666, + 0.014735255390405655, + 0.029331179335713387, + 0.008341542445123196, + 0.04816820099949837, + 0.012228043749928474, + 0.007673601619899273, + 0.005320609547197819, + 0.009900843724608421, + 0.010518189519643784, + 0.00882507860660553, + 0.010730410926043987, + 0.02058872953057289, + 0.016231635585427284, + 0.009902813471853733, + 0.007873708382248878, + 0.011899915523827076, + 0.009216729551553726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.020586160942912102 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015455812215805 + ], + "opposite_logprob_tokens": [ + -1.7015455961227417 + ], + "logprob_reward": -0.2015455812215805, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015455812215805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1576797217130661, + 0.05876676365733147, + 0.05864748731255531, + 0.06176787614822388, + 0.05072762817144394, + 0.07868018001317978, + 0.11362601816654205, + 0.04516938328742981, + 0.30985313653945923, + 0.06508185714483261 + ], + "credit_map": [ + 0.046688199043273926, + 0.020599380135536194, + 0.009063152596354485, + 0.01033154595643282, + 0.02122238650918007, + 0.01363383885473013, + 0.010014832019805908, + 0.026126381009817123, + 0.010680449195206165, + 0.005036171991378069, + 0.013408794067800045, + 0.00556252570822835, + 0.006094049662351608, + 0.00526482192799449, + 0.00471888342872262, + 0.008001063950359821, + 0.004909573122859001, + 0.005801919847726822, + 0.010723278857767582, + 0.006082741543650627, + 0.00999438390135765, + 0.006472443696111441, + 0.008109075948596, + 0.006554066203534603, + 0.007173147052526474, + 0.009203044697642326, + 0.010890357196331024, + 0.008014524355530739, + 0.007048647850751877, + 0.00758607080206275, + 0.00514665013179183, + 0.006705433130264282, + 0.006285746116191149, + 0.006772376596927643, + 0.0064536575227975845, + 0.004854061175137758, + 0.008852820843458176, + 0.005442544352263212, + 0.005354297813028097, + 0.0067121172323822975, + 0.010333109647035599, + 0.008558820933103561, + 0.01083796750754118, + 0.010248237289488316, + 0.014031761325895786, + 0.005971647333353758, + 0.007866021245718002, + 0.010832604020833969, + 0.010576234199106693, + 0.009708813391625881, + 0.028082244098186493, + 0.011717853136360645, + 0.010942358523607254, + 0.009173721075057983, + 0.010753043927252293, + 0.022671740502119064, + 0.006393929943442345, + 0.004642830695956945, + 0.009285720065236092, + 0.006062673404812813, + 0.00408745976164937, + 0.004410233814269304, + 0.004389932379126549, + 0.005896601360291243, + 0.0747031643986702, + 0.0222723800688982, + 0.11126785725355148, + 0.023952383548021317, + 0.023684579879045486, + 0.013718868605792522, + 0.02220705710351467, + 0.018046824261546135, + 0.00646719615906477, + 0.008004098199307919, + 0.014407530426979065, + 0.010270974598824978, + 0.006046561524271965, + 0.005280252546072006, + 0.008814231492578983, + 0.005791008938103914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.2015455812215805 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide-angle shot captures a full-body view of a clump of iron-arrow bamboo, rendered entirely in the stark, high-contrast black-and-white ink style of Charles Burns. The camera remains completely still, rendering the scene like a horrifying panel from a graphic novel. Amidst the pure, inky void, the bamboo vines don't just wither; they twist into unnatural, harrowing shapes, their surfaces riddled with grotesque, tumorous protrusions and deep, shadowy pits that defy their natural form. Leaves curl inward like the dehydrated, arthritic fingers of a corpse, each a sickly, flat sheet of white. True to the designated style, each leaf vein is meticulously rendered as a sharp black line against the sickly white, emphasizing the fragile, lifeless texture. The harsh lighting, with no discernible source, creates a world of absolute binary opposition—no grays, no soft gradients, only solid black shapes and stark, unforgiving white surfaces. This high-contrast aesthetic creates a deeply unsettling atmosphere of bio-horror and decay, transforming the plants into giant, static specimens of botanical mutation.", + "seed": 25253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_3/index_413_seed_25253_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07092523481696844, + "vlm_logprob": -0.07092523481696844, + "vlm_yes_no_margin": 3.199999976158142, + "vlm_token_credit": -0.07092523481696844, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07414809614419937 + ], + "opposite_logprob_tokens": [ + -2.699148178100586 + ], + "logprob_reward": -0.07414809614419937, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07414809614419937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36342599987983704, + 0.35331422090530396, + 0.06062948331236839, + 0.06349248439073563, + 0.02683480829000473, + 0.021993521600961685, + 0.024108638986945152, + 0.019004622474312782, + 0.05116158723831177, + 0.01603461243212223 + ], + "credit_map": [ + 0.04999972879886627, + 0.05016736686229706, + 0.022931739687919617, + 0.032897479832172394, + 0.040585532784461975, + 0.022466985508799553, + 0.06912478804588318, + 0.07525239884853363, + 0.032826945185661316, + 0.019017131999135017, + 0.082241490483284, + 0.13168644905090332, + 0.017254069447517395, + 0.01923372969031334, + 0.01723472774028778, + 0.03381970152258873, + 0.005475240759551525, + 0.009000571444630623, + 0.012478763237595558, + 0.007921737618744373, + 0.006505614146590233, + 0.005821634083986282, + 0.007324846927076578, + 0.006101078353822231, + 0.004191041458398104, + 0.004425529856234789, + 0.015846125781536102, + 0.021449904888868332, + 0.003397434949874878, + 0.0027755589690059423, + 0.006174615118652582, + 0.005232269410043955, + 0.0028891663532704115, + 0.0028418758884072304, + 0.007581944111734629, + 0.002426964696496725, + 0.0020523546263575554, + 0.002421557903289795, + 0.0034886121284216642, + 0.00313233258202672, + 0.001902588875964284, + 0.0021790368482470512, + 0.002116019604727626, + 0.002580085536465049, + 0.002710426226258278, + 0.0017425146652385592, + 0.005753038451075554, + 0.0030098124407231808, + 0.002242659218609333, + 0.0036938630510121584, + 0.0037674603518098593, + 0.002356837037950754, + 0.0026329706888645887, + 0.0028973855078220367, + 0.003756863996386528, + 0.002760599832981825, + 0.0020890822634100914, + 0.002669954439625144, + 0.0022319480776786804, + 0.0040733302012085915, + 0.002024163492023945, + 0.001632926519960165, + 0.0016995895421132445, + 0.002583628986030817, + 0.0060965013690292835, + 0.005469487514346838, + 0.013846796937286854, + 0.012729262001812458, + 0.005919063929468393, + 0.0024108390789479017, + 0.002388653112575412, + 0.0023009839933365583, + 0.0019000080646947026, + 0.0017710295505821705, + 0.0019093586597591639, + 0.0028747795149683952, + 0.0016956226900219917, + 0.0017966983141377568, + 0.0015537618892267346, + 0.002533355727791786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.07414809614419937 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129913330078125 + ], + "opposite_logprob_tokens": [ + -2.2379913330078125 + ], + "logprob_reward": -0.1129913330078125, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129913330078125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12672540545463562, + 0.16897983849048615, + 0.1630258858203888, + 0.1321525275707245, + 0.12687960267066956, + 0.04104076698422432, + 0.031830936670303345, + 0.04371938481926918, + 0.060680028051137924, + 0.10496557503938675 + ], + "credit_map": [ + 0.020189622417092323, + 0.017669064924120903, + 0.013506432995200157, + 0.012583594769239426, + 0.010221259668469429, + 0.009367048740386963, + 0.016545873135328293, + 0.02664251998066902, + 0.01548786647617817, + 0.012108834460377693, + 0.04318869858980179, + 0.04822979494929314, + 0.011410696431994438, + 0.011100109666585922, + 0.015512704849243164, + 0.011941144242882729, + 0.013261905871331692, + 0.016248509287834167, + 0.0341399610042572, + 0.051532454788684845, + 0.010450930334627628, + 0.011226312257349491, + 0.011302885599434376, + 0.014862941578030586, + 0.011574584990739822, + 0.009621369652450085, + 0.031676292419433594, + 0.04489629715681076, + 0.0071791429072618484, + 0.007389696780592203, + 0.009151443839073181, + 0.010663711465895176, + 0.010396374389529228, + 0.01148364320397377, + 0.01455435249954462, + 0.03341692313551903, + 0.00905212014913559, + 0.00936880987137556, + 0.01319907046854496, + 0.02540830709040165, + 0.004120516590774059, + 0.005916782189160585, + 0.004836112726479769, + 0.005764830391854048, + 0.003754136385396123, + 0.00349411997012794, + 0.004089077468961477, + 0.009065194055438042, + 0.004203688818961382, + 0.0054849800653755665, + 0.004631904885172844, + 0.0030984152108430862, + 0.003069231752306223, + 0.0025607235729694366, + 0.005405600648373365, + 0.0033763928804546595, + 0.006311056669801474, + 0.006232307758182287, + 0.01174697931855917, + 0.0037046088837087154, + 0.005133237224072218, + 0.0036437278613448143, + 0.0032202161382883787, + 0.0037272549234330654, + 0.004522045608609915, + 0.004841784946620464, + 0.014821239747107029, + 0.006097780074924231, + 0.019772378727793694, + 0.0035615582019090652, + 0.003096685977652669, + 0.003966554068028927, + 0.00715441070497036, + 0.007248288951814175, + 0.02255946956574917, + 0.03756467252969742, + 0.011111137457191944, + 0.007466472219675779, + 0.005433667916804552, + 0.006427456624805927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.1129913330078125 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100012592971325 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.004100012592971325, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100012592971325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28538310527801514, + 0.21482546627521515, + 0.09348657727241516, + 0.0523061640560627, + 0.08694428205490112, + 0.05651106685400009, + 0.018230289220809937, + 0.029196761548519135, + 0.13347411155700684, + 0.029642188921570778 + ], + "credit_map": [ + 0.0399504154920578, + 0.037863511592149734, + 0.017052914947271347, + 0.023402778431773186, + 0.032280176877975464, + 0.017797274515032768, + 0.06414037942886353, + 0.05289563909173012, + 0.016959963366389275, + 0.011690514162182808, + 0.05639141798019409, + 0.07397516071796417, + 0.012628045864403248, + 0.011453568004071712, + 0.014409354887902737, + 0.017317429184913635, + 0.008561874739825726, + 0.009900158271193504, + 0.016944605857133865, + 0.012755591422319412, + 0.010107280686497688, + 0.006973233539611101, + 0.020762499421834946, + 0.007481333799660206, + 0.0039270720444619656, + 0.004261596594005823, + 0.015058264136314392, + 0.015460258349776268, + 0.0025673212949186563, + 0.002903967397287488, + 0.0039145746268332005, + 0.004213106818497181, + 0.005993874277919531, + 0.006116236560046673, + 0.013500127010047436, + 0.02345103584229946, + 0.007244944106787443, + 0.005098987836390734, + 0.011928881518542767, + 0.01361020002514124, + 0.005028713494539261, + 0.005549725145101547, + 0.01384375523775816, + 0.008496175520122051, + 0.005971115548163652, + 0.005341000854969025, + 0.006778647191822529, + 0.005501936189830303, + 0.0014939921675249934, + 0.0025294891092926264, + 0.0029553407803177834, + 0.0018635188462212682, + 0.0023507275618612766, + 0.0020826430991292, + 0.002518917666748166, + 0.0024356599897146225, + 0.004198417067527771, + 0.0030651669949293137, + 0.005121744703501463, + 0.00546950101852417, + 0.0024562315084040165, + 0.0031522063072770834, + 0.0028325545135885477, + 0.002900937804952264, + 0.008019370958209038, + 0.006914394907653332, + 0.020442839711904526, + 0.027794107794761658, + 0.050225723534822464, + 0.006223215721547604, + 0.005897355731576681, + 0.007957108318805695, + 0.0025354514364153147, + 0.002310963114723563, + 0.007536803372204304, + 0.008756235241889954, + 0.002523850416764617, + 0.001826729392632842, + 0.0018836288945749402, + 0.0022685276344418526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.004100012592971325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058534324169159 + ], + "opposite_logprob_tokens": [ + -3.895585298538208 + ], + "logprob_reward": -0.02058534324169159, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058534324169159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13831713795661926, + 0.19127093255519867, + 0.07456360012292862, + 0.051722683012485504, + 0.14501233398914337, + 0.04936077445745468, + 0.03446870297193527, + 0.06217446178197861, + 0.18185290694236755, + 0.07125647366046906 + ], + "credit_map": [ + 0.013451330363750458, + 0.012351376935839653, + 0.01713421754539013, + 0.013720926828682423, + 0.018581952899694443, + 0.009334986098110676, + 0.03911236673593521, + 0.014629996381700039, + 0.01636521890759468, + 0.010694775730371475, + 0.04528586566448212, + 0.06719454377889633, + 0.010888307355344296, + 0.013322727754712105, + 0.011283726431429386, + 0.01623578369617462, + 0.010187148116528988, + 0.013594881631433964, + 0.010611314326524734, + 0.009991715662181377, + 0.007745968643575907, + 0.0056526693515479565, + 0.009895098395645618, + 0.006884817499667406, + 0.0048431348986923695, + 0.004510532598942518, + 0.018070843070745468, + 0.010786620900034904, + 0.0028825413901358843, + 0.0033042074646800756, + 0.003984295763075352, + 0.0033405106514692307, + 0.012808123603463173, + 0.021280696615576744, + 0.015826068818569183, + 0.030076144263148308, + 0.012277266941964626, + 0.009538925252854824, + 0.021208809688687325, + 0.021996324881911278, + 0.004713116213679314, + 0.005675389431416988, + 0.00855528749525547, + 0.0068413130939006805, + 0.005359923001378775, + 0.005095271393656731, + 0.006827860604971647, + 0.0062926160171628, + 0.003641261486336589, + 0.003541602985933423, + 0.005078136455267668, + 0.004549638368189335, + 0.0035667349584400654, + 0.0029831412248313427, + 0.0038932461757212877, + 0.007214944344013929, + 0.00691758980974555, + 0.006954883225262165, + 0.016201747581362724, + 0.011749839410185814, + 0.005563732236623764, + 0.0054145376197993755, + 0.004551177844405174, + 0.0048209610395133495, + 0.01363266073167324, + 0.013454833999276161, + 0.053345877677202225, + 0.013018959201872349, + 0.061372850090265274, + 0.008591270074248314, + 0.006841263733804226, + 0.011595210991799831, + 0.0049468218348920345, + 0.00511451531201601, + 0.02044876664876938, + 0.025299955159425735, + 0.006174186710268259, + 0.0029410459101200104, + 0.002708564978092909, + 0.0036226219963282347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 4 + ], + "selected_reward": -0.02058534324169159 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a clump of iron-arrow bamboo shown in a full-body view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered in stark black-and-white ink with no gray gradients?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the bamboo vines display grotesque, tumorous protrusions and deep shadowy pits?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the leaves depicted as curled inward, flat white sheets resembling arthritic fingers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and wide-angle, maintaining a fixed position throughout?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280138909816742 + ], + "opposite_logprob_tokens": [ + -2.017801284790039 + ], + "logprob_reward": -0.14280138909816742, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14280138909816742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23173609375953674, + 0.26236817240715027, + 0.0712432786822319, + 0.040324173867702484, + 0.03254307061433792, + 0.04125149920582771, + 0.01842544414103031, + 0.02601785957813263, + 0.1283470094203949, + 0.14774344861507416 + ], + "credit_map": [ + 0.026327289640903473, + 0.03373048081994057, + 0.02035738155245781, + 0.029530387371778488, + 0.032971885055303574, + 0.02526676282286644, + 0.028994254767894745, + 0.034557633101940155, + 0.021168177947402, + 0.013186321593821049, + 0.07288245856761932, + 0.09288366138935089, + 0.01379932276904583, + 0.013867845758795738, + 0.01596529595553875, + 0.01861509121954441, + 0.005621761083602905, + 0.008472924120724201, + 0.012499959208071232, + 0.018763253465294838, + 0.005964852869510651, + 0.005542477127164602, + 0.008484404534101486, + 0.005893642082810402, + 0.0031199632212519646, + 0.003692583180963993, + 0.008977283723652363, + 0.006944131571799517, + 0.0032813732977956533, + 0.0026667281053960323, + 0.007822802290320396, + 0.003819307778030634, + 0.0034197564236819744, + 0.0032219463028013706, + 0.008860504254698753, + 0.0034801678266376257, + 0.0028329750057309866, + 0.0026609732303768396, + 0.005288236308842897, + 0.002778512891381979, + 0.0038380410987883806, + 0.004587800242006779, + 0.009121648035943508, + 0.003558556782081723, + 0.0035757399164140224, + 0.00279083289206028, + 0.007396330125629902, + 0.006382550112903118, + 0.0020779292099177837, + 0.0025271582417190075, + 0.0025532920844852924, + 0.00242184498347342, + 0.0017629367066547275, + 0.0016230143373832107, + 0.002739305840805173, + 0.0027199634350836277, + 0.0031424700282514095, + 0.002636654768139124, + 0.005023397039622068, + 0.00388162094168365, + 0.0031167941633611917, + 0.002606503665447235, + 0.0029445583932101727, + 0.00266586197540164, + 0.011382680386304855, + 0.012014302425086498, + 0.046364493668079376, + 0.009108860976994038, + 0.031027451157569885, + 0.005408701486885548, + 0.005267363507300615, + 0.007773146964609623, + 0.010314985178411007, + 0.010997731238603592, + 0.046235281974077225, + 0.04839324951171875, + 0.011159186251461506, + 0.006162088830024004, + 0.006316222716122866, + 0.008164693601429462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.14280138909816742 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_4/index_tensor([1069, 1069])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_4/index_tensor([1069, 1069])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..527a2d1725ac41bf75c7283e8bcb94101ea393fb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_4/index_tensor([1069, 1069])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, shot from a horizontal angle, is rendered in stunning high-contrast black and white. Against a backdrop of boundless, pure black, three to five dumplings, each bulging in shape and form, silently and leisurely descend from above in an elegant slow-motion sequence. A strong, sharply defined overhead light illuminates them with pinpoint precision, sculpting their rounded contours and setting their bright white skins against the depths of the background. Encased in delicate, almost translucent dough wrappers, the hand-kneaded folds at the edges cast subtle shadows under the strong light, highlighting their handcrafted quality. As they fall, the dumplings gently twirl and tumble in the air, performing a silent ballet in a weightless environment. The scene, imbued with a minimalist aesthetic, creates an atmosphere of tranquility, concentration, and poetry, elevating the simple form of food into a dynamic work of art.", + "seed": 10963, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_4/index_1069_seed_10963_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.6849089376628399, + "vlm_logprob": -0.6849089376628399, + "vlm_yes_no_margin": 1.025, + "vlm_token_credit": -0.6849089376628399, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07088316977024078 + ], + "opposite_logprob_tokens": [ + -2.945883274078369 + ], + "logprob_reward": -0.07088316977024078, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07088316977024078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021741684526205063, + 0.03556233271956444, + 0.8214417099952698, + 0.027945702895522118, + 0.037232864648103714, + 0.02073485776782036, + 0.01121385395526886, + 0.0051131038926541805, + 0.00882764346897602, + 0.010186238214373589 + ], + "credit_map": [ + 0.0026934188790619373, + 0.0016617454821243882, + 0.0026466697454452515, + 0.0017639918951317668, + 0.0022852756083011627, + 0.0056983609683811665, + 0.0019324802560731769, + 0.0030597432050853968, + 0.0009609202388674021, + 0.0009303974220529199, + 0.0020019994117319584, + 0.0007369366358034313, + 0.015967106446623802, + 0.012484748847782612, + 0.0009520911844447255, + 0.0015281338710337877, + 0.0374140739440918, + 0.008265888318419456, + 0.03219543769955635, + 0.011303826235234737, + 0.14074987173080444, + 0.2897793650627136, + 0.07516321539878845, + 0.22657008469104767, + 0.000612337957136333, + 0.0004901930806227028, + 0.0008451368776150048, + 0.0006398065597750247, + 0.0013834253186360002, + 0.009740379638969898, + 0.012424219399690628, + 0.0018102050526067615, + 0.001996136037632823, + 0.0008635486010462046, + 0.01400772761553526, + 0.0009587574750185013, + 0.003156570252031088, + 0.012873916886746883, + 0.0012276358902454376, + 0.0021485735196620226, + 0.0011834093602374196, + 0.001298009417951107, + 0.005196128040552139, + 0.0024098290596157312, + 0.0012077555293217301, + 0.0023775878362357616, + 0.0015486275078728795, + 0.005513512529432774, + 0.0007837796001695096, + 0.0009294943301938474, + 0.002460940508171916, + 0.0009225050453096628, + 0.0007582498947158456, + 0.0010857523884624243, + 0.0012021501315757632, + 0.0030709838028997183, + 0.0003485306224320084, + 0.0003923187905456871, + 0.0003613222506828606, + 0.0006274730549193919, + 0.0003445885959081352, + 0.001405557501129806, + 0.0004365368513390422, + 0.0011967766331508756, + 0.0010694083757698536, + 0.0013344399631023407, + 0.0009127594530582428, + 0.0006026597693562508, + 0.0010143164545297623, + 0.0015676275361329317, + 0.0009058554423972964, + 0.0014205771731212735, + 0.0017688210355117917, + 0.001878709881566465, + 0.0007766446215100586, + 0.0007338345749303699, + 0.0007227597525343299, + 0.001583467936143279, + 0.0009859452256932855, + 0.00173605652526021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.07088316977024078 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134193420410156 + ], + "opposite_logprob_tokens": [ + -0.31341928243637085 + ], + "logprob_reward": -1.3134193420410156, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134193420410156, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07889871299266815, + 0.03954546898603439, + 0.3591567277908325, + 0.3358694016933441, + 0.05562722682952881, + 0.01327106636017561, + 0.012498179450631142, + 0.016213295981287956, + 0.06848429888486862, + 0.02043558843433857 + ], + "credit_map": [ + 0.005621740594506264, + 0.002425538841634989, + 0.00842330139130354, + 0.0023367854300886393, + 0.003585566533729434, + 0.023747628554701805, + 0.0036456244997680187, + 0.02911253273487091, + 0.0012582434574142098, + 0.001663434086367488, + 0.0016636289656162262, + 0.0016103932866826653, + 0.014635203406214714, + 0.01573367416858673, + 0.0012111098039895296, + 0.001769782043993473, + 0.016874175518751144, + 0.003613273613154888, + 0.014093978330492973, + 0.004919377155601978, + 0.061445947736501694, + 0.1271759569644928, + 0.03177747130393982, + 0.09925656765699387, + 0.006350227165967226, + 0.004474034998565912, + 0.007974171079695225, + 0.006893329322338104, + 0.015524208545684814, + 0.13155710697174072, + 0.1479392796754837, + 0.015157060697674751, + 0.002040785737335682, + 0.0015445295721292496, + 0.012691515497863293, + 0.0014454303309321404, + 0.003908198326826096, + 0.023212654516100883, + 0.0017394209280610085, + 0.009044692851603031, + 0.0008305698283948004, + 0.0012061807792633772, + 0.0033141691237688065, + 0.0013075810857117176, + 0.0008563327137380838, + 0.001600605552084744, + 0.000944133847951889, + 0.003211492905393243, + 0.0007532653980888426, + 0.0005823936080560088, + 0.00384697038680315, + 0.0006666856352239847, + 0.0005304886144585907, + 0.0009908892679959536, + 0.000987578765489161, + 0.0041399081237614155, + 0.0014338471228256822, + 0.0015092553803697228, + 0.0013981801457703114, + 0.0018653757870197296, + 0.001579264528118074, + 0.003782395040616393, + 0.0017790313577279449, + 0.0028659470845013857, + 0.008012672886252403, + 0.007548808120191097, + 0.004664949607104063, + 0.003035909030586481, + 0.005433748476207256, + 0.01593979261815548, + 0.005063241813331842, + 0.018785180523991585, + 0.003943168558180332, + 0.003008810570463538, + 0.0013238477986305952, + 0.0013093235902488232, + 0.0010445243678987026, + 0.00378540251404047, + 0.0009984871139749885, + 0.005022023804485798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -1.3134193420410156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371308371424675 + ], + "opposite_logprob_tokens": [ + -3.4087131023406982 + ], + "logprob_reward": -0.03371308371424675, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371308371424675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.002542913891375065, + 0.009563340805470943, + 0.9574677348136902, + 0.016644928604364395, + 0.006206165999174118, + 0.004654488060623407, + 0.0008476590155623853, + 0.0005316184833645821, + 0.0006706307758577168, + 0.0008705719956196845 + ], + "credit_map": [ + 0.00030592724215239286, + 0.00015567858645226806, + 0.0004872059798799455, + 0.00014807646221015602, + 0.000207624543691054, + 0.000422691140556708, + 0.00018109897791873664, + 0.0006346107111312449, + 0.00020109732577111572, + 0.0002844209666363895, + 0.00030287806293927133, + 0.0003254264302086085, + 0.0033406505826860666, + 0.003885105485096574, + 0.0004003442300017923, + 0.0008234168635681272, + 0.04404696822166443, + 0.009707021526992321, + 0.03754889592528343, + 0.01328104455024004, + 0.16357600688934326, + 0.3374479413032532, + 0.08757032454013824, + 0.2642894387245178, + 0.00031119288178160787, + 0.00022550890571437776, + 0.0003715754428412765, + 0.0003457684942986816, + 0.0007854521973058581, + 0.0067944759503006935, + 0.007216446567326784, + 0.000594508193898946, + 0.00024079630384221673, + 0.00021344792912714183, + 0.0005786623223684728, + 0.00018881574214901775, + 0.000631756498478353, + 0.0023909311275929213, + 0.00026345832156948745, + 0.0016982967499643564, + 0.00016729027265682817, + 0.0002422456891508773, + 0.001401666086167097, + 0.00028068400570191443, + 0.000139769705128856, + 0.000436618021922186, + 0.00033935403916984797, + 0.0016468598041683435, + 0.00011228200310142711, + 9.788173338165507e-05, + 0.00012872782826889306, + 9.049235814018175e-05, + 7.550504233222455e-05, + 0.00010675772500690073, + 0.00010888272663578391, + 0.00012712954776361585, + 4.567179348669015e-05, + 5.2362629503477365e-05, + 5.640235394821502e-05, + 7.302853919100016e-05, + 5.191486707190052e-05, + 8.694260759511963e-05, + 8.629298099549487e-05, + 7.900266791693866e-05, + 8.701467595528811e-05, + 9.890885121421888e-05, + 6.0984399169683456e-05, + 5.855529525433667e-05, + 0.00013635598588734865, + 8.512930071447045e-05, + 7.269551133504137e-05, + 7.098671630956233e-05, + 0.0001592938497196883, + 0.00010307411866961047, + 7.327820640057325e-05, + 5.683322888216935e-05, + 7.114215986803174e-05, + 0.00013911337009631097, + 8.425224950769916e-05, + 0.00018358469242230058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.03371308371424675 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8048416376113892 + ], + "opposite_logprob_tokens": [ + -0.17984165251255035 + ], + "logprob_reward": -1.8048416376113892, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048416376113892, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1548498123884201, + 0.12926723062992096, + 0.25284332036972046, + 0.20309698581695557, + 0.07756135612726212, + 0.032180994749069214, + 0.017222898080945015, + 0.02774127759039402, + 0.04382263496518135, + 0.061413466930389404 + ], + "credit_map": [ + 0.011257902719080448, + 0.003877749666571617, + 0.019237687811255455, + 0.006688735447824001, + 0.01026199758052826, + 0.04099614545702934, + 0.009425141848623753, + 0.0531044527888298, + 0.0029895782936364412, + 0.0029689271468669176, + 0.006158299278467894, + 0.0016936318716034293, + 0.059119079262018204, + 0.04823945835232735, + 0.002858540043234825, + 0.005239715334028006, + 0.011650769971311092, + 0.0025846941862255335, + 0.00993516854941845, + 0.0035191518254578114, + 0.04337247088551521, + 0.08930311352014542, + 0.022315531969070435, + 0.07016243040561676, + 0.004598262254148722, + 0.0023352946154773235, + 0.004603807348757982, + 0.004169609397649765, + 0.00954858586192131, + 0.08747827261686325, + 0.07898741215467453, + 0.011375746689736843, + 0.0031954713631421328, + 0.001996637787669897, + 0.020825505256652832, + 0.0018334194319322705, + 0.007259433623403311, + 0.03128965571522713, + 0.00277638528496027, + 0.008384850807487965, + 0.0033171558752655983, + 0.003143587615340948, + 0.005856059025973082, + 0.003619299503043294, + 0.0022342924494296312, + 0.0041109733283519745, + 0.0035582494456321, + 0.006341379135847092, + 0.0013607284054160118, + 0.0017363459337502718, + 0.0028458351735025644, + 0.001651681843213737, + 0.001336361630819738, + 0.0025535726454108953, + 0.002490083687007427, + 0.003248288994655013, + 0.00262586516328156, + 0.0017751591512933373, + 0.001980297965928912, + 0.0030451547354459763, + 0.0018284996040165424, + 0.006073283962905407, + 0.003347087884321809, + 0.007065927609801292, + 0.004508549813181162, + 0.004849029239267111, + 0.0036237689200788736, + 0.0025174058973789215, + 0.006030216347426176, + 0.00816869456321001, + 0.004429113119840622, + 0.009695859625935555, + 0.011061213910579681, + 0.009521769359707832, + 0.0037791586946696043, + 0.004731651861220598, + 0.003585828235372901, + 0.011295904405415058, + 0.0031606690026819706, + 0.014277270063757896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -1.8048416376113892 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20168745517730713 + ], + "opposite_logprob_tokens": [ + -1.7016874551773071 + ], + "logprob_reward": -0.20168745517730713, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20168745517730713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02918838895857334, + 0.08621122688055038, + 0.6819039583206177, + 0.0845755785703659, + 0.05997525155544281, + 0.010487879626452923, + 0.01685793697834015, + 0.003244033083319664, + 0.019916728138923645, + 0.007638986222445965 + ], + "credit_map": [ + 0.0023321318440139294, + 0.0011206383351236582, + 0.0023050655145198107, + 0.0012144428910687566, + 0.001591799664311111, + 0.009436960332095623, + 0.001679057371802628, + 0.009508296847343445, + 0.0018573052948340774, + 0.0018529292428866029, + 0.003538146847859025, + 0.0019006881630048156, + 0.03749094530940056, + 0.034915052354335785, + 0.002523919567465782, + 0.002132250228896737, + 0.03111962042748928, + 0.006943427491933107, + 0.026614567264914513, + 0.009341096505522728, + 0.11596796661615372, + 0.2406797856092453, + 0.06211794912815094, + 0.18911957740783691, + 0.0015559158055111766, + 0.0010197256924584508, + 0.0019372064853087068, + 0.0016517980257049203, + 0.004118291661143303, + 0.03422744199633598, + 0.036997485905885696, + 0.0030677157919853926, + 0.0028523090295493603, + 0.0015685781836509705, + 0.020487014204263687, + 0.0015143458731472492, + 0.004912486765533686, + 0.02388044260442257, + 0.0021213917061686516, + 0.0026386897079646587, + 0.000733029970433563, + 0.0009039390715770423, + 0.0019157121423631907, + 0.0016392046818509698, + 0.000809202843811363, + 0.0013013818534091115, + 0.0011491236509755254, + 0.0020362865179777145, + 0.0011433875188231468, + 0.0010611657053232193, + 0.0039015442598611116, + 0.0011813793098554015, + 0.001128515345044434, + 0.002475690795108676, + 0.0018027016194537282, + 0.004163552541285753, + 0.0005017765215598047, + 0.00021635988377965987, + 0.00038330123061314225, + 0.0003743427514564246, + 0.00029662830638699234, + 0.0005192879470996559, + 0.0003483635955490172, + 0.0006039729923941195, + 0.001959514571353793, + 0.0029971806798130274, + 0.0017471122555434704, + 0.0019168062135577202, + 0.002289373427629471, + 0.0037521780468523502, + 0.0016173557378351688, + 0.0036372097674757242, + 0.0010088401613757014, + 0.0013055611634626985, + 0.000756950001232326, + 0.0006346859154291451, + 0.0008225515484809875, + 0.001185969216749072, + 0.0005969227058812976, + 0.0013275061501190066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.20168745517730713 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a horizontal angle, is rendered in stunning high-contrast black and white. Against a backdrop of boundless, pure black, three to five dumplings, each bulging in shape and form, silently and leisurely descend from above in an elegant slow-motion sequence. A strong, sharply defined overhead light illuminates them with pinpoint precision, sculpting their rounded contours and setting their bright white skins against the depths of the background. Encased in delicate, almost translucent dough wrappers, the hand-kneaded folds at the edges cast subtle shadows under the strong light, highlighting their handcrafted quality. As they fall, the dumplings gently twirl and tumble in the air, performing a silent ballet in a weightless environment. The scene, imbued with a minimalist aesthetic, creates an atmosphere of tranquility, concentration, and poetry, elevating the simple form of food into a dynamic work of art.", + "seed": 10964, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_4/index_1069_seed_10964_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.48291480392217634, + "vlm_logprob": -0.48291480392217634, + "vlm_yes_no_margin": 1.425, + "vlm_token_credit": -0.48291480392217634, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1902579665184021 + ], + "opposite_logprob_tokens": [ + -1.8152580261230469 + ], + "logprob_reward": -0.1902579665184021, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1902579665184021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06388441473245621, + 0.01851152814924717, + 0.016742462292313576, + 0.022247903048992157, + 0.02645866572856903, + 0.026445481926202774, + 0.0662456527352333, + 0.11863839626312256, + 0.5956771969795227, + 0.0451483279466629 + ], + "credit_map": [ + 0.007816066034138203, + 0.00565921887755394, + 0.016009457409381866, + 0.003945188131183386, + 0.0074803754687309265, + 0.011691879481077194, + 0.003790420014411211, + 0.007491800934076309, + 0.0016908677062019706, + 0.002648932160809636, + 0.0027080464642494917, + 0.0031688448507338762, + 0.0020738737657666206, + 0.0020377645269036293, + 0.0022475668229162693, + 0.001935629639774561, + 0.0013310527428984642, + 0.002128619235008955, + 0.003850797191262245, + 0.001976082567125559, + 0.002965987427160144, + 0.0015529559459537268, + 0.0017764937365427613, + 0.0011604717001318932, + 0.001506210770457983, + 0.0025223696138709784, + 0.003338672686368227, + 0.0032064926344901323, + 0.003934939857572317, + 0.0044702026061713696, + 0.002056325785815716, + 0.0012126871151849627, + 0.001901081996038556, + 0.0035996115766465664, + 0.005293424241244793, + 0.00437525799497962, + 0.0027277928311377764, + 0.0031360387802124023, + 0.0035183983854949474, + 0.001907056663185358, + 0.0016931784339249134, + 0.004622486419975758, + 0.004137500189244747, + 0.0028505921363830566, + 0.0024726251140236855, + 0.005790055263787508, + 0.0026054116897284985, + 0.0022736291866749525, + 0.004734396934509277, + 0.009374825283885002, + 0.009645174257457256, + 0.007116352207958698, + 0.014534459449350834, + 0.009367215447127819, + 0.005901732016354799, + 0.005571490619331598, + 0.01016928255558014, + 0.011221249587833881, + 0.02161555364727974, + 0.018076693639159203, + 0.012909244745969772, + 0.01360431220382452, + 0.019726954400539398, + 0.011315100826323032, + 0.030179884284734726, + 0.015560980886220932, + 0.23886512219905853, + 0.015873443335294724, + 0.017697831615805626, + 0.23991698026657104, + 0.01570780761539936, + 0.02187507227063179, + 0.0035548009909689426, + 0.003961554728448391, + 0.013303045183420181, + 0.002267783507704735, + 0.0033395886421203613, + 0.010734369046986103, + 0.004294934682548046, + 0.003692242782562971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.1902579665184021 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994044721126556 + ], + "opposite_logprob_tokens": [ + -1.8049404621124268 + ], + "logprob_reward": -0.17994044721126556, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994044721126556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23108500242233276, + 0.11565472930669785, + 0.14270319044589996, + 0.14548300206661224, + 0.0398222878575325, + 0.02744237519800663, + 0.08943638205528259, + 0.04350464046001434, + 0.1305854320526123, + 0.03428295999765396 + ], + "credit_map": [ + 0.02509370818734169, + 0.02387094870209694, + 0.04300990328192711, + 0.018726132810115814, + 0.03536485880613327, + 0.0372995063662529, + 0.024570992216467857, + 0.02314893715083599, + 0.009300119243562222, + 0.016431689262390137, + 0.03142275661230087, + 0.007687102537602186, + 0.008756655268371105, + 0.025033818557858467, + 0.006426712963730097, + 0.01059587299823761, + 0.012106861919164658, + 0.0161297507584095, + 0.03652268275618553, + 0.019264062866568565, + 0.022779205814003944, + 0.014714585617184639, + 0.011170271784067154, + 0.010015761479735374, + 0.014300246722996235, + 0.019862551242113113, + 0.026501478627324104, + 0.02264804020524025, + 0.017309358343482018, + 0.021728238090872765, + 0.012681601569056511, + 0.010451486334204674, + 0.002960805082693696, + 0.005493943113833666, + 0.004743785131722689, + 0.0060623520985245705, + 0.005218313075602055, + 0.004851273261010647, + 0.0076210820116102695, + 0.002870730357244611, + 0.0017733200220391154, + 0.004686141852289438, + 0.004326069261878729, + 0.004740301053971052, + 0.002745317993685603, + 0.0035221586003899574, + 0.003284212900325656, + 0.002364850603044033, + 0.006030193530023098, + 0.01591246947646141, + 0.012213661335408688, + 0.008721692487597466, + 0.018612761050462723, + 0.013949809595942497, + 0.006745145656168461, + 0.00725064892321825, + 0.0035367871169000864, + 0.004494627472013235, + 0.007506657391786575, + 0.006092064548283815, + 0.0050210487097501755, + 0.005891556851565838, + 0.007186663802713156, + 0.0037752294447273016, + 0.00714450515806675, + 0.00537899462506175, + 0.043850529938936234, + 0.0046407547779381275, + 0.0073772333562374115, + 0.04855291172862053, + 0.00564555125311017, + 0.007994935847818851, + 0.004108259920030832, + 0.0028230533935129642, + 0.004476596601307392, + 0.0025241735856980085, + 0.004641068633645773, + 0.0056989723816514015, + 0.006938477046787739, + 0.0030723540112376213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.17994044721126556 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033709973096847534 + ], + "opposite_logprob_tokens": [ + -3.40871000289917 + ], + "logprob_reward": -0.033709973096847534, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033709973096847534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10328712314367294, + 0.01386439148336649, + 0.01817653328180313, + 0.05210565775632858, + 0.017929092049598694, + 0.033721644431352615, + 0.05147459730505943, + 0.06450504064559937, + 0.6050143837928772, + 0.039921581745147705 + ], + "credit_map": [ + 0.014122617430984974, + 0.009248918853700161, + 0.023798935115337372, + 0.006897012237459421, + 0.009505445137619972, + 0.019589951261878014, + 0.007422990631312132, + 0.012701245956122875, + 0.0011909583117812872, + 0.0014560014242306352, + 0.0032084863632917404, + 0.0017318945610895753, + 0.0015116469003260136, + 0.001743101398460567, + 0.001842453726567328, + 0.0011798485647886992, + 0.001311667961999774, + 0.0027932936791330576, + 0.0023805410601198673, + 0.0026821568608283997, + 0.0032476019114255905, + 0.0024540594313293695, + 0.0018587898230180144, + 0.0014484220882877707, + 0.0041082170791924, + 0.008524429053068161, + 0.005924616474658251, + 0.006155176088213921, + 0.01086726039648056, + 0.0075377607718110085, + 0.0050820112228393555, + 0.003906187368556857, + 0.0014368625124916434, + 0.0023469654843211174, + 0.0033816248178482056, + 0.0029006307013332844, + 0.0020000895019620657, + 0.001997824991121888, + 0.0022850900422781706, + 0.0015800031833350658, + 0.0026109744794666767, + 0.007385683711618185, + 0.0033483151346445084, + 0.004383002873510122, + 0.0024158605374395847, + 0.0076751988381147385, + 0.002888453658670187, + 0.003014154499396682, + 0.004410342313349247, + 0.007926492020487785, + 0.006367775145918131, + 0.005847633350640535, + 0.008456126786768436, + 0.00793529860675335, + 0.005124981515109539, + 0.00540594756603241, + 0.0069466400891542435, + 0.008320936933159828, + 0.008638506755232811, + 0.008822329342365265, + 0.012901789508759975, + 0.007224408909678459, + 0.006217717658728361, + 0.005432717502117157, + 0.028880860656499863, + 0.010306143201887608, + 0.24503834545612335, + 0.013815870508551598, + 0.01525602675974369, + 0.2558288276195526, + 0.012506912462413311, + 0.02338137850165367, + 0.004586666822433472, + 0.0031658175867050886, + 0.007953263819217682, + 0.002268747193738818, + 0.004068155772984028, + 0.009732631966471672, + 0.0045818062499165535, + 0.0035644914023578167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.033709973096847534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031840950250626 + ], + "opposite_logprob_tokens": [ + -2.350318431854248 + ], + "logprob_reward": -0.10031840950250626, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031840950250626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42051342129707336, + 0.17288298904895782, + 0.13026338815689087, + 0.13460080325603485, + 0.03280356526374817, + 0.02319534309208393, + 0.025433216243982315, + 0.01534202229231596, + 0.0284862257540226, + 0.016479091718792915 + ], + "credit_map": [ + 0.05683707073330879, + 0.04113670811057091, + 0.08906377106904984, + 0.032970067113637924, + 0.04522741213440895, + 0.07342934608459473, + 0.0295285414904356, + 0.05232048034667969, + 0.013883940875530243, + 0.028440631926059723, + 0.04087366908788681, + 0.011680105701088905, + 0.016219012439250946, + 0.03790890425443649, + 0.008034816943109035, + 0.015841888263821602, + 0.01007675752043724, + 0.017471399158239365, + 0.024562425911426544, + 0.016557415947318077, + 0.02513919770717621, + 0.015548479743301868, + 0.012097790837287903, + 0.008809911087155342, + 0.011960219591856003, + 0.017028862610459328, + 0.023795107379555702, + 0.019572606310248375, + 0.01650054194033146, + 0.024769248440861702, + 0.011298520490527153, + 0.009675695560872555, + 0.002655658172443509, + 0.0037566134706139565, + 0.007265667896717787, + 0.00574727775529027, + 0.0024019803386181593, + 0.005184778943657875, + 0.0032253090757876635, + 0.0025662793777883053, + 0.0017428367864340544, + 0.00556746032088995, + 0.003404255723580718, + 0.00325569836422801, + 0.0017869619186967611, + 0.0037268968299031258, + 0.002053789095953107, + 0.0016574420733377337, + 0.0018438854021951556, + 0.0046485066413879395, + 0.003708301577717066, + 0.0024028366897255182, + 0.005399482790380716, + 0.0035354676656425, + 0.0018220447236672044, + 0.002072688890621066, + 0.0016428623348474503, + 0.0014136810787022114, + 0.002224844880402088, + 0.0018901353469118476, + 0.002942449878901243, + 0.002088189125061035, + 0.0016115190228447318, + 0.001528339460492134, + 0.0033242914360016584, + 0.0026715167332440615, + 0.004356267862021923, + 0.002269101794809103, + 0.004066102672368288, + 0.005719845648854971, + 0.0023526025470346212, + 0.0037264947313815355, + 0.001441769185476005, + 0.0009368657483719289, + 0.004052568227052689, + 0.001046338933520019, + 0.0017139815026894212, + 0.0035912971943616867, + 0.0023472211323678493, + 0.001349048689007759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.10031840950250626 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.9103472232818604 + ], + "opposite_logprob_tokens": [ + -0.16034728288650513 + ], + "logprob_reward": -1.9103472232818604, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.9103472232818604, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05734316259622574, + 0.02676670253276825, + 0.024943822994828224, + 0.030683336779475212, + 0.014528139494359493, + 0.04743187502026558, + 0.10988323390483856, + 0.1122417151927948, + 0.4741422235965729, + 0.10203582793474197 + ], + "credit_map": [ + 0.007283102720975876, + 0.005177619867026806, + 0.012427237816154957, + 0.003633505431935191, + 0.0077971238642930984, + 0.009892436675727367, + 0.004066415596753359, + 0.007065721787512302, + 0.0028589204885065556, + 0.004301964771002531, + 0.00447689788416028, + 0.002239775378257036, + 0.0028394032269716263, + 0.005540424957871437, + 0.002195627661421895, + 0.00231368956156075, + 0.002084296429529786, + 0.0032784007489681244, + 0.005587155930697918, + 0.0038247788324952126, + 0.0034481515176594257, + 0.0027537508867681026, + 0.00225169793702662, + 0.001715589314699173, + 0.002555328654125333, + 0.0062868595123291016, + 0.0035984416026622057, + 0.003933134488761425, + 0.005230749025940895, + 0.004046423826366663, + 0.00328624970279634, + 0.0017461497336626053, + 0.001317395712248981, + 0.002249047625809908, + 0.0021515442058444023, + 0.002547486452385783, + 0.0015061571029946208, + 0.001852845074608922, + 0.0017825426766648889, + 0.0011211209930479527, + 0.003690792480483651, + 0.009348961524665356, + 0.010671514086425304, + 0.005483685992658138, + 0.0022265578154474497, + 0.009257867932319641, + 0.003165421076118946, + 0.0035870715510100126, + 0.008767264895141125, + 0.011589148081839085, + 0.017067573964595795, + 0.013430282473564148, + 0.020404668524861336, + 0.017228590324521065, + 0.011267675086855888, + 0.010128019377589226, + 0.01224823109805584, + 0.007124142721295357, + 0.023586519062519073, + 0.01203239057213068, + 0.017468366771936417, + 0.019140832126140594, + 0.009053574874997139, + 0.011587665416300297, + 0.023227475583553314, + 0.0076796975918114185, + 0.1914849877357483, + 0.011189993470907211, + 0.01213749311864376, + 0.19947591423988342, + 0.009632899425923824, + 0.019313747063279152, + 0.012192840687930584, + 0.007273008581250906, + 0.02264316938817501, + 0.006191933527588844, + 0.008445504121482372, + 0.025185441598296165, + 0.01143653690814972, + 0.00866739172488451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -1.9103472232818604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static medium shot, shot from a horizontal angle, is rendered in stunning high-contrast black and white. Against a backdrop of boundless, pure black, three to five dumplings, each bulging in shape and form, silently and leisurely descend from above in an elegant slow-motion sequence. A strong, sharply defined overhead light illuminates them with pinpoint precision, sculpting their rounded contours and setting their bright white skins against the depths of the background. Encased in delicate, almost translucent dough wrappers, the hand-kneaded folds at the edges cast subtle shadows under the strong light, highlighting their handcrafted quality. As they fall, the dumplings gently twirl and tumble in the air, performing a silent ballet in a weightless environment. The scene, imbued with a minimalist aesthetic, creates an atmosphere of tranquility, concentration, and poetry, elevating the simple form of food into a dynamic work of art.", + "seed": 10965, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_4/index_1069_seed_10965_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14032230265438556, + "vlm_logprob": -0.14032230265438556, + "vlm_yes_no_margin": 2.450000011920929, + "vlm_token_credit": -0.14032230265438556, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07098822295665741 + ], + "opposite_logprob_tokens": [ + -2.945988178253174 + ], + "logprob_reward": -0.07098822295665741, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07098822295665741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05688401311635971, + 0.04543444886803627, + 0.033466488122940063, + 0.025131680071353912, + 0.01754186302423477, + 0.044198088347911835, + 0.05025651678442955, + 0.27520516514778137, + 0.019089149311184883, + 0.43279266357421875 + ], + "credit_map": [ + 0.007802726235240698, + 0.007704535964876413, + 0.010014383122324944, + 0.003955752123147249, + 0.007450244855135679, + 0.009654974564909935, + 0.004842015448957682, + 0.0054593756794929504, + 0.0053014690056443214, + 0.005018521565943956, + 0.010141829960048199, + 0.004568297881633043, + 0.004241016693413258, + 0.006720834411680698, + 0.004712845664471388, + 0.004729628097265959, + 0.00400279276072979, + 0.004754319787025452, + 0.004912137985229492, + 0.004690379370003939, + 0.00291971187107265, + 0.004130474757403135, + 0.004078085999935865, + 0.003978581167757511, + 0.0032543290872126818, + 0.0029515824280679226, + 0.0044571757316589355, + 0.0024416851811110973, + 0.0029224322643131018, + 0.00419828062877059, + 0.002861735178157687, + 0.0020444553811103106, + 0.0033739565405994654, + 0.00115790287964046, + 0.0028930550906807184, + 0.0010482053039595485, + 0.0024987156502902508, + 0.0022780681028962135, + 0.002093367977067828, + 0.0021985892672091722, + 0.005482085980474949, + 0.003248256631195545, + 0.009252241812646389, + 0.004962773062288761, + 0.0042536016553640366, + 0.005382876377552748, + 0.0036900753621011972, + 0.007926173508167267, + 0.0039864215068519115, + 0.003860476426780224, + 0.011628332547843456, + 0.004785122815519571, + 0.0027060965076088905, + 0.014214293099939823, + 0.003017324721440673, + 0.006058442406356335, + 0.013147714547812939, + 0.006316734943538904, + 0.10746757686138153, + 0.007511322386562824, + 0.005932703614234924, + 0.11471565812826157, + 0.005372544750571251, + 0.014740857295691967, + 0.0023271350655704737, + 0.00138570845592767, + 0.0027905083261430264, + 0.0017453012987971306, + 0.0027136479038745165, + 0.003593537723645568, + 0.0013942319201305509, + 0.0031390765216201544, + 0.02208743430674076, + 0.004372614901512861, + 0.19001217186450958, + 0.011770134791731834, + 0.00941113755106926, + 0.16601355373859406, + 0.007801448460668325, + 0.021324120461940765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.07098822295665741 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990827560424805 + ], + "opposite_logprob_tokens": [ + -1.804908275604248 + ], + "logprob_reward": -0.17990827560424805, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990827560424805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1348707675933838, + 0.11204934120178223, + 0.050915561616420746, + 0.03482532501220703, + 0.0351264514029026, + 0.011702005751430988, + 0.103399857878685, + 0.1506783366203308, + 0.015264524146914482, + 0.35116779804229736 + ], + "credit_map": [ + 0.022386357188224792, + 0.02293374575674534, + 0.024600328877568245, + 0.0079347537830472, + 0.010775509290397167, + 0.02678401954472065, + 0.009116238914430141, + 0.01033981703221798, + 0.013751933351159096, + 0.016801128163933754, + 0.020008768886327744, + 0.008935411460697651, + 0.010087456554174423, + 0.015718698501586914, + 0.01252907607704401, + 0.014216872863471508, + 0.005795125383883715, + 0.00629026535898447, + 0.009027034044265747, + 0.005718359258025885, + 0.004672923590987921, + 0.007513695862144232, + 0.006288510747253895, + 0.005609650630503893, + 0.005713651422411203, + 0.0027265953831374645, + 0.0064911493100225925, + 0.0033782885875552893, + 0.003540981560945511, + 0.00513761630281806, + 0.002932167612016201, + 0.004904870875179768, + 0.006212988868355751, + 0.0027604398783296347, + 0.00524223642423749, + 0.001754962489940226, + 0.004809667356312275, + 0.00446406239643693, + 0.0031227932777255774, + 0.006759300362318754, + 0.0018569895764812827, + 0.000924046675208956, + 0.0018600036855787039, + 0.0012470460496842861, + 0.0014619843568652868, + 0.0017847440904006362, + 0.001462785992771387, + 0.0011044054990634322, + 0.010619855485856533, + 0.003562997095286846, + 0.03804473578929901, + 0.005211889743804932, + 0.0035937551874667406, + 0.026866115629673004, + 0.004250836092978716, + 0.011249673552811146, + 0.006847519427537918, + 0.0041363537311553955, + 0.0595603808760643, + 0.0035173811484128237, + 0.00315107312053442, + 0.06314743310213089, + 0.0022631220053881407, + 0.008055073209106922, + 0.001620373222976923, + 0.0008525209850631654, + 0.0027343761175870895, + 0.0010230769403278828, + 0.0009465018520131707, + 0.004461980424821377, + 0.000948169850744307, + 0.0026775242295116186, + 0.018950091674923897, + 0.00479569286108017, + 0.15020088851451874, + 0.010619254782795906, + 0.010298416949808598, + 0.13287939131259918, + 0.0073990244418382645, + 0.016025040298700333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.17990827560424805 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306739568710327 + ], + "opposite_logprob_tokens": [ + -3.168067455291748 + ], + "logprob_reward": -0.04306739568710327, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306739568710327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10091855376958847, + 0.08876040577888489, + 0.05468559265136719, + 0.04037533327937126, + 0.048630356788635254, + 0.12676431238651276, + 0.0807291641831398, + 0.2786763608455658, + 0.04618403688073158, + 0.1342758983373642 + ], + "credit_map": [ + 0.012382914312183857, + 0.014525352977216244, + 0.01568451337516308, + 0.00982911977916956, + 0.012684935703873634, + 0.014792882837355137, + 0.011921020224690437, + 0.009097817353904247, + 0.009698184207081795, + 0.014231717213988304, + 0.014597855508327484, + 0.007111016195267439, + 0.012545708566904068, + 0.011219805106520653, + 0.010338758118450642, + 0.009017353877425194, + 0.007008029613643885, + 0.007771956734359264, + 0.008193364366889, + 0.007211722899228334, + 0.004536045715212822, + 0.007458608131855726, + 0.006078663747757673, + 0.006427199579775333, + 0.0077501265332102776, + 0.002922292100265622, + 0.008901416324079037, + 0.003123407019302249, + 0.0035889139398932457, + 0.007430933881551027, + 0.0027977891732007265, + 0.003860456869006157, + 0.008084847591817379, + 0.0035187567118555307, + 0.007490635849535465, + 0.0035987563896924257, + 0.006509664002805948, + 0.005833413917571306, + 0.005827745888382196, + 0.0077665382996201515, + 0.013707268051803112, + 0.007557434495538473, + 0.038627903908491135, + 0.008081848733127117, + 0.007959512993693352, + 0.017700180411338806, + 0.007069014012813568, + 0.026061149314045906, + 0.008573008701205254, + 0.005094890948385, + 0.028219202533364296, + 0.006451189052313566, + 0.003256658324971795, + 0.018076715990900993, + 0.0038351225666701794, + 0.007222374901175499, + 0.012625844217836857, + 0.007511884905397892, + 0.10147778689861298, + 0.007955294102430344, + 0.006988647393882275, + 0.1149703860282898, + 0.00700633367523551, + 0.020140184089541435, + 0.005842363927513361, + 0.002892578486353159, + 0.00831362884491682, + 0.003404998453333974, + 0.006779841613024473, + 0.01036535482853651, + 0.0026582987047731876, + 0.005926973652094603, + 0.010641434229910374, + 0.00413494510576129, + 0.0498473085463047, + 0.005097848363220692, + 0.007701112888753414, + 0.030434636399149895, + 0.006363660097122192, + 0.020054945722222328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.04306739568710327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058417536318302 + ], + "opposite_logprob_tokens": [ + -3.8955841064453125 + ], + "logprob_reward": -0.02058417536318302, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058417536318302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06088288873434067, + 0.19265635311603546, + 0.037767887115478516, + 0.057284604758024216, + 0.061959706246852875, + 0.01732580177485943, + 0.15043185651302338, + 0.23228533565998077, + 0.013858860358595848, + 0.17554661631584167 + ], + "credit_map": [ + 0.009705663658678532, + 0.007651609368622303, + 0.014723734930157661, + 0.0037973939906805754, + 0.004948358051478863, + 0.009614691138267517, + 0.004059726372361183, + 0.006381717044860125, + 0.028458837419748306, + 0.0225503109395504, + 0.03959612920880318, + 0.013583398424088955, + 0.015919141471385956, + 0.029946789145469666, + 0.01807759515941143, + 0.02452417090535164, + 0.004715509247034788, + 0.006226927507668734, + 0.005681230220943689, + 0.0038808821700513363, + 0.002898424630984664, + 0.0065803565084934235, + 0.003258688375353813, + 0.004525871016085148, + 0.008759510703384876, + 0.004644046071916819, + 0.012386804446578026, + 0.004626371432095766, + 0.005412178114056587, + 0.009809338487684727, + 0.004586979281157255, + 0.007059385068714619, + 0.012384012341499329, + 0.003935922868549824, + 0.009407978504896164, + 0.004203550051897764, + 0.006003386341035366, + 0.009421703405678272, + 0.0047506727278232574, + 0.011852488853037357, + 0.0033709777053445578, + 0.0015424111625179648, + 0.0027580747846513987, + 0.0020799138583242893, + 0.0019479364855214953, + 0.002349718939512968, + 0.0013726995093747973, + 0.0019040703773498535, + 0.012575078755617142, + 0.0040565417148172855, + 0.059935834258794785, + 0.004900594707578421, + 0.0036354553885757923, + 0.0524480976164341, + 0.003873508656397462, + 0.00900675356388092, + 0.010736025869846344, + 0.010479357093572617, + 0.08556248992681503, + 0.005926957819610834, + 0.006382369436323643, + 0.09321559965610504, + 0.004345494322478771, + 0.015637055039405823, + 0.0018150069518014789, + 0.0009946380741894245, + 0.003525092266499996, + 0.0007479538908228278, + 0.0011248543160036206, + 0.003835556097328663, + 0.0007621667464263737, + 0.0010535928886383772, + 0.007394071668386459, + 0.0018096447456628084, + 0.0677318200469017, + 0.004527767188847065, + 0.00387826981022954, + 0.07555777579545975, + 0.00326238083653152, + 0.011384888552129269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.02058417536318302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870634436607361 + ], + "opposite_logprob_tokens": [ + -1.1370635032653809 + ], + "logprob_reward": -0.3870634436607361, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3870634436607361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10036497563123703, + 0.10824473947286606, + 0.035704970359802246, + 0.04389077052474022, + 0.03693714737892151, + 0.04026670753955841, + 0.0993892252445221, + 0.36823827028274536, + 0.06831633299589157, + 0.09864692389965057 + ], + "credit_map": [ + 0.011078252457082272, + 0.01659827120602131, + 0.01659221015870571, + 0.007828354835510254, + 0.012357567436993122, + 0.01690066047012806, + 0.00923222303390503, + 0.009777436032891273, + 0.014646204188466072, + 0.013766438700258732, + 0.022450704127550125, + 0.008275308646261692, + 0.008615046739578247, + 0.01597408391535282, + 0.011312202550470829, + 0.013204752467572689, + 0.00476965494453907, + 0.0034336845856159925, + 0.00452039809897542, + 0.004110730718821287, + 0.005843701772391796, + 0.004787974525243044, + 0.0043875607661902905, + 0.003851264948025346, + 0.006981323007494211, + 0.0030505533795803785, + 0.0077888257801532745, + 0.004435938782989979, + 0.004632239229977131, + 0.007640990428626537, + 0.0036434833891689777, + 0.005717413034290075, + 0.005542055703699589, + 0.0029465574771165848, + 0.005818489473313093, + 0.0036091217771172523, + 0.004883310757577419, + 0.005087253637611866, + 0.003755547571927309, + 0.005294808652251959, + 0.007974128238856792, + 0.0032199432607740164, + 0.00699353963136673, + 0.004299146588891745, + 0.003501707687973976, + 0.005905400030314922, + 0.0038924661930650473, + 0.004480375442653894, + 0.008745378814637661, + 0.0027655570302158594, + 0.02395421452820301, + 0.0034994171001017094, + 0.002449219347909093, + 0.036563631147146225, + 0.0024438153486698866, + 0.018967987969517708, + 0.018838508054614067, + 0.011821393854916096, + 0.15119677782058716, + 0.00961776077747345, + 0.008700510486960411, + 0.14766938984394073, + 0.00740272318944335, + 0.012991206720471382, + 0.006615051068365574, + 0.0036457867827266455, + 0.016745805740356445, + 0.003057456109672785, + 0.0032756351865828037, + 0.023168232291936874, + 0.003887658007442951, + 0.007920709438621998, + 0.011223329231142998, + 0.010025804862380028, + 0.01431217510253191, + 0.010632630437612534, + 0.012644092552363873, + 0.01681992970407009, + 0.011289154179394245, + 0.011699806898832321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.3870634436607361 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, shot from a horizontal angle, is rendered in stunning high-contrast black and white. Against a backdrop of boundless, pure black, three to five dumplings, each bulging in shape and form, silently and leisurely descend from above in an elegant slow-motion sequence. A strong, sharply defined overhead light illuminates them with pinpoint precision, sculpting their rounded contours and setting their bright white skins against the depths of the background. Encased in delicate, almost translucent dough wrappers, the hand-kneaded folds at the edges cast subtle shadows under the strong light, highlighting their handcrafted quality. As they fall, the dumplings gently twirl and tumble in the air, performing a silent ballet in a weightless environment. The scene, imbued with a minimalist aesthetic, creates an atmosphere of tranquility, concentration, and poetry, elevating the simple form of food into a dynamic work of art.", + "seed": 10966, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_4/index_1069_seed_10966_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.31145391464233396, + "vlm_logprob": -0.31145391464233396, + "vlm_yes_no_margin": 1.6, + "vlm_token_credit": -0.31145391464233396, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08862824738025665 + ], + "opposite_logprob_tokens": [ + -2.5886282920837402 + ], + "logprob_reward": -0.08862824738025665, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08862824738025665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2113845944404602, + 0.6403064131736755, + 0.01583518274128437, + 0.0140219796448946, + 0.023323552682995796, + 0.007063170429319143, + 0.008328784257173538, + 0.02276196889579296, + 0.031561270356178284, + 0.0254130270332098 + ], + "credit_map": [ + 0.017151309177279472, + 0.029337681829929352, + 0.044602442532777786, + 0.020875686779618263, + 0.02585427276790142, + 0.03165685012936592, + 0.023836694657802582, + 0.01806965097784996, + 0.029191358014941216, + 0.008252502419054508, + 0.2517940402030945, + 0.012615572661161423, + 0.012097158469259739, + 0.2826061546802521, + 0.008004769682884216, + 0.035744860768318176, + 0.001557617331854999, + 0.0023484693374484777, + 0.002966943196952343, + 0.0013486124807968736, + 0.002746267942711711, + 0.0020339838229119778, + 0.0013333344832062721, + 0.0014999554259702563, + 0.0009729386656545103, + 0.002228242577984929, + 0.003365197451785207, + 0.001328264712356031, + 0.0014624163741245866, + 0.002070246497169137, + 0.0015856365207582712, + 0.0010090370196849108, + 0.0012230384163558483, + 0.004430277738720179, + 0.003533909097313881, + 0.002275870880112052, + 0.0032459613867104053, + 0.004273361526429653, + 0.002496446017175913, + 0.001844688318669796, + 0.0007378109730780125, + 0.0008684425847604871, + 0.0013016253942623734, + 0.001198174781166017, + 0.0007360095623880625, + 0.000866957358084619, + 0.0005730248522013426, + 0.0007811248069629073, + 0.0007701488211750984, + 0.0010734446113929152, + 0.0009430413483642042, + 0.001608771039173007, + 0.0011870886664837599, + 0.0006596362218260765, + 0.001222577877342701, + 0.0008640762534923851, + 0.002425262937322259, + 0.002640391234308481, + 0.00380144314840436, + 0.004836332518607378, + 0.0015238100895658135, + 0.00171279301866889, + 0.0036708335392177105, + 0.00215110182762146, + 0.003439531894400716, + 0.004397395998239517, + 0.004018530249595642, + 0.00375998392701149, + 0.004427039530128241, + 0.004182571545243263, + 0.0036761814262717962, + 0.003660035552456975, + 0.0023180656135082245, + 0.0022916404996067286, + 0.005768172442913055, + 0.0026602211873978376, + 0.0020614885725080967, + 0.003000633092597127, + 0.003601444186642766, + 0.0037113609723746777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.08862824738025665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990708351135254 + ], + "opposite_logprob_tokens": [ + -1.8049070835113525 + ], + "logprob_reward": -0.17990708351135254, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990708351135254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3028090000152588, + 0.10511280596256256, + 0.264109343290329, + 0.051099058240652084, + 0.03782186284661293, + 0.04665055871009827, + 0.02872660756111145, + 0.033277206122875214, + 0.049775972962379456, + 0.08061757683753967 + ], + "credit_map": [ + 0.027646856382489204, + 0.010221393778920174, + 0.09170496463775635, + 0.015010800212621689, + 0.01880851574242115, + 0.08752723783254623, + 0.011344468221068382, + 0.04054475203156471, + 0.005998666863888502, + 0.005658093374222517, + 0.030504291877150536, + 0.005382970906794071, + 0.006420383229851723, + 0.0348711721599102, + 0.006668343674391508, + 0.009608880616724491, + 0.017040586099028587, + 0.030585812404751778, + 0.09723613411188126, + 0.012423884123563766, + 0.029939135536551476, + 0.05052836239337921, + 0.009722325019538403, + 0.01663312315940857, + 0.0036171632818877697, + 0.009994393214583397, + 0.006993948481976986, + 0.007496871519833803, + 0.0061498889699578285, + 0.007706803269684315, + 0.005191718693822622, + 0.0039482731372118, + 0.0038790341932326555, + 0.005808594170957804, + 0.005250141955912113, + 0.005891165230423212, + 0.003927541896700859, + 0.005910488776862621, + 0.0035374737344682217, + 0.0036174235865473747, + 0.004595697391778231, + 0.005378163419663906, + 0.011166541837155819, + 0.007254377938807011, + 0.00395815959200263, + 0.0061374325305223465, + 0.0039492417126894, + 0.0042109438218176365, + 0.0028550133574754, + 0.003350113285705447, + 0.0039941975846886635, + 0.005381446331739426, + 0.003934529609978199, + 0.0022966444958001375, + 0.004116639960557222, + 0.002798024332150817, + 0.0029453046154230833, + 0.0033554527908563614, + 0.005463623441755772, + 0.007677887100726366, + 0.0034537555184215307, + 0.0027118916623294353, + 0.004820922389626503, + 0.0028483665082603693, + 0.005670073442161083, + 0.005640798714011908, + 0.006532432045787573, + 0.006278481334447861, + 0.005681497976183891, + 0.006907141767442226, + 0.006817014422267675, + 0.006248532794415951, + 0.0054747555404901505, + 0.004605116788297892, + 0.026737436652183533, + 0.0045824614353477955, + 0.004001951310783625, + 0.022969013080000877, + 0.0072128428146243095, + 0.0050340015441179276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.17990708351135254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988118529319763 + ], + "opposite_logprob_tokens": [ + -1.80488121509552 + ], + "logprob_reward": -0.17988118529319763, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988118529319763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034809671342372894, + 0.7846940755844116, + 0.031895533204078674, + 0.023377718403935432, + 0.018484190106391907, + 0.014129027724266052, + 0.009247925132513046, + 0.015408019535243511, + 0.0469733290374279, + 0.020980454981327057 + ], + "credit_map": [ + 0.0032549547031521797, + 0.003612023778259754, + 0.005381537601351738, + 0.0038016601465642452, + 0.00393354007974267, + 0.005332370288670063, + 0.004539028275758028, + 0.004954555071890354, + 0.035804297775030136, + 0.009970740415155888, + 0.31180235743522644, + 0.015001701191067696, + 0.01469087041914463, + 0.3465311527252197, + 0.009577294811606407, + 0.041315700858831406, + 0.003077700035646558, + 0.004417142830789089, + 0.00828769151121378, + 0.0021606003865599632, + 0.0038563860580325127, + 0.004064691253006458, + 0.0030639003962278366, + 0.0029674223624169827, + 0.002112496178597212, + 0.004111281596124172, + 0.0030315814074128866, + 0.0036475355736911297, + 0.002382908947765827, + 0.003564310260117054, + 0.002798794535920024, + 0.0017288110684603453, + 0.0011449964949861169, + 0.003027387196198106, + 0.0026541065890341997, + 0.0022796362172812223, + 0.0027830174658447504, + 0.003059083828702569, + 0.002020440064370632, + 0.0015155208529904485, + 0.0015175885055214167, + 0.0016606374410912395, + 0.0018823634600266814, + 0.002627805108204484, + 0.0014335338491946459, + 0.0020946504082530737, + 0.0016376290004700422, + 0.0012748193694278598, + 0.0008622246095910668, + 0.0010652876226231456, + 0.001319886650890112, + 0.001547282561659813, + 0.0013040262274444103, + 0.0011932805646210909, + 0.0011561868013814092, + 0.0007997495704330504, + 0.0015049774665385485, + 0.0013227564049884677, + 0.0026525473222136497, + 0.0034030398819595575, + 0.0014004834229126573, + 0.0013161099050194025, + 0.002284268382936716, + 0.0015238382620736957, + 0.007221714593470097, + 0.004401902202516794, + 0.006145840045064688, + 0.005149355623871088, + 0.004708531312644482, + 0.00479615293443203, + 0.0062448675744235516, + 0.008304966613650322, + 0.0019609781447798014, + 0.0016278745606541634, + 0.005527499131858349, + 0.001856441842392087, + 0.001394774066284299, + 0.002620865125209093, + 0.0025389783550053835, + 0.003453044453635812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.17988118529319763 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495734512805939 + ], + "opposite_logprob_tokens": [ + -2.929957389831543 + ], + "logprob_reward": -0.05495734512805939, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495734512805939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032241735607385635, + 0.7927693724632263, + 0.026064958423376083, + 0.02381661906838417, + 0.05797562003135681, + 0.0330696739256382, + 0.014271066524088383, + 0.006389521062374115, + 0.008403507992625237, + 0.004997990094125271 + ], + "credit_map": [ + 0.004133518785238266, + 0.0022587350104004145, + 0.0053229257464408875, + 0.0038132949266582727, + 0.0029246441554278135, + 0.006635741330683231, + 0.0025372144300490618, + 0.004615661688148975, + 0.035846300423145294, + 0.010197284631431103, + 0.3160262107849121, + 0.015494143590331078, + 0.014902759343385696, + 0.3501574397087097, + 0.009831229224801064, + 0.040314022451639175, + 0.0018834901275113225, + 0.0034444350749254227, + 0.01000058464705944, + 0.0012524824123829603, + 0.0024277702905237675, + 0.0039789858274161816, + 0.0011036002542823553, + 0.001973611768335104, + 0.0019432406406849623, + 0.0042303744703531265, + 0.0028405268676579, + 0.003720371052622795, + 0.0029500836972147226, + 0.004749756772071123, + 0.0017942807171493769, + 0.0015879836864769459, + 0.003235932905226946, + 0.012194350361824036, + 0.0084456792101264, + 0.006383589934557676, + 0.0055858981795609, + 0.015006103552877903, + 0.0030146881472319365, + 0.004109374713152647, + 0.0024750682059675455, + 0.0038976448122411966, + 0.006244204938411713, + 0.007257984019815922, + 0.0023392552975565195, + 0.004207805264741182, + 0.0034089265391230583, + 0.0032387850806117058, + 0.0014965293230488896, + 0.0017543197609484196, + 0.002385108731687069, + 0.002328143222257495, + 0.0012849981430917978, + 0.00130087963771075, + 0.002171285916119814, + 0.0015498013235628605, + 0.0008369117858819664, + 0.0008591506630182266, + 0.0008940050029195845, + 0.0009829734917730093, + 0.0005698165623471141, + 0.0007488479022867978, + 0.0009033763199113309, + 0.0005944392178207636, + 0.0009585593943484128, + 0.0007514010067097843, + 0.001173354801721871, + 0.0011772818397730589, + 0.0009825442684814334, + 0.0012775149662047625, + 0.001060300157405436, + 0.0010225519072264433, + 0.0006968154339119792, + 0.000531562662217766, + 0.0006719650700688362, + 0.0003690223675221205, + 0.0005440119421109557, + 0.0007754871039651334, + 0.0004873384314123541, + 0.0009217872284352779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.05495734512805939 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are there between three and five dumplings visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dumplings descend from above in a slow-motion sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a boundless, pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium shot from a horizontal angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538957118988037 + ], + "opposite_logprob_tokens": [ + -0.4288956820964813 + ], + "logprob_reward": -1.0538957118988037, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538957118988037, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09535335749387741, + 0.2770235538482666, + 0.3227664530277252, + 0.06208887696266174, + 0.09279244393110275, + 0.022513676434755325, + 0.033323563635349274, + 0.018263496458530426, + 0.04020103067159653, + 0.035673581063747406 + ], + "credit_map": [ + 0.010278522036969662, + 0.0068595753982663155, + 0.022442035377025604, + 0.007667970843613148, + 0.009215490892529488, + 0.016778234392404556, + 0.006023832596838474, + 0.016087688505649567, + 0.013960562646389008, + 0.0042622992768883705, + 0.11111237108707428, + 0.006161130033433437, + 0.006147843785583973, + 0.1190207228064537, + 0.003941551316529512, + 0.012417035177350044, + 0.023942099884152412, + 0.038729459047317505, + 0.1296277642250061, + 0.014283526688814163, + 0.039235737174749374, + 0.03984541445970535, + 0.013210954144597054, + 0.02389146387577057, + 0.004883487243205309, + 0.011418575420975685, + 0.011903680860996246, + 0.007863527163863182, + 0.0062325866892933846, + 0.006703462917357683, + 0.007972260005772114, + 0.005111295264214277, + 0.004243830218911171, + 0.023333873599767685, + 0.012164604850113392, + 0.008136079646646976, + 0.007426162716001272, + 0.0251899603754282, + 0.005086068995296955, + 0.007211865857243538, + 0.0019637583754956722, + 0.0030718278139829636, + 0.004426244180649519, + 0.0023695984855294228, + 0.0028690299950540066, + 0.003315605456009507, + 0.0020088129676878452, + 0.0024887993931770325, + 0.003589965868741274, + 0.005112756043672562, + 0.006299520377069712, + 0.0035984618589282036, + 0.004055223427712917, + 0.0024911744985729456, + 0.003223270643502474, + 0.004953186959028244, + 0.0022323643788695335, + 0.002521072281524539, + 0.0030175549909472466, + 0.003581831231713295, + 0.0013849438400939107, + 0.001227712957188487, + 0.0025118053890764713, + 0.0017862102249637246, + 0.00717042526230216, + 0.00319503853097558, + 0.005073183681815863, + 0.004288364667445421, + 0.00283380807377398, + 0.004365474916994572, + 0.004974010866135359, + 0.0083007188513875, + 0.0029384687077254057, + 0.0017963870195671916, + 0.012255745939910412, + 0.0018010882195085287, + 0.0012345543364062905, + 0.010457462631165981, + 0.0017374390736222267, + 0.003452431643381715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -1.0538957118988037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_5/index_tensor([6669, 6669])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_5/index_tensor([6669, 6669])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..86dcca82adb09479497616a2d879d27f0b1b5165 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_5/index_tensor([6669, 6669])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot depicts the solitary figure of a young man, Dojima, standing motionless on a deserted train platform at night. He has dark, slightly unkempt hair, his face tilted slightly upward, an expression of quiet, melancholic wonder. He wears a simple charcoal-gray hoodie over a T-shirt, his hands tucked deep into the pockets of his dark jeans. The platform itself, made of weathered concrete and empty save for a dark wooden bench, is bathed in the cold, sterile light of a single overhead lamp, which casts long, sharp shadows across the surrounding darkness. Above this silent scene, the universe is in awe-inspiring motion. In a dramatic time-lapse effect, the stars, rather than fixed points, become long, dazzling trails of light, arcing and swirling across the deep indigo canvas of the night sky. The visual contrast between the static, contemplative young man and the frenetic, silent ballet of the universe creates a powerful atmosphere of serene solitude and awe, capturing a single, poignant moment of human existence against the backdrop of cosmic eternity.", + "seed": 7260, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_5/index_6669_seed_7260_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06997113209217787, + "vlm_logprob": -0.06997113209217787, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.06997113209217787, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051806893199682236 + ], + "opposite_logprob_tokens": [ + -3.051806926727295 + ], + "logprob_reward": -0.051806893199682236, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.051806893199682236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27533164620399475, + 0.16912241280078888, + 0.09402378648519516, + 0.0678970217704773, + 0.05032740905880928, + 0.06423231959342957, + 0.04953307658433914, + 0.07815610617399216, + 0.06944417953491211, + 0.08193198591470718 + ], + "credit_map": [ + 0.03560469672083855, + 0.03860552981495857, + 0.027262285351753235, + 0.03834807500243187, + 0.02830464020371437, + 0.029706906527280807, + 0.038200728595256805, + 0.039298783987760544, + 0.01696559600532055, + 0.022793056443333626, + 0.019997816532850266, + 0.015835430473089218, + 0.021802639588713646, + 0.020008908584713936, + 0.02426038682460785, + 0.02745858021080494, + 0.01341572217643261, + 0.015800515189766884, + 0.012206874787807465, + 0.012750791385769844, + 0.007618463132530451, + 0.010758464224636555, + 0.009811930358409882, + 0.01166102197021246, + 0.008193565532565117, + 0.010499962605535984, + 0.007886803708970547, + 0.0070329937152564526, + 0.009839166887104511, + 0.009098141454160213, + 0.008633390069007874, + 0.0067129977978765965, + 0.005639629438519478, + 0.009524333290755749, + 0.00636745011433959, + 0.006925878114998341, + 0.0044914004392921925, + 0.006932887714356184, + 0.005978825036436319, + 0.0044670067727565765, + 0.005863676778972149, + 0.015272540040314198, + 0.006289568264037371, + 0.008599268272519112, + 0.004780652467161417, + 0.009061207994818687, + 0.005336339119821787, + 0.009029071778059006, + 0.0042125154286623, + 0.006925176829099655, + 0.004448248539119959, + 0.006535240449011326, + 0.00565497949719429, + 0.007101770024746656, + 0.007108152844011784, + 0.007546992041170597, + 0.007453834172338247, + 0.010182433761656284, + 0.0060338410548865795, + 0.007140605244785547, + 0.015775460749864578, + 0.009521927684545517, + 0.010723124258220196, + 0.01132488064467907, + 0.008654765784740448, + 0.013383212499320507, + 0.007767706178128719, + 0.005478848237544298, + 0.006953380536288023, + 0.010936195962131023, + 0.007218178361654282, + 0.009051896631717682, + 0.00947907567024231, + 0.01092150341719389, + 0.00812547467648983, + 0.006819542031735182, + 0.010980996303260326, + 0.012712622061371803, + 0.011408940888941288, + 0.011483831331133842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.051806893199682236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214574359357357 + ], + "opposite_logprob_tokens": [ + -4.264214515686035 + ], + "logprob_reward": -0.014214574359357357, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214574359357357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2140907645225525, + 0.18617293238639832, + 0.09512345492839813, + 0.07593666762113571, + 0.051012054085731506, + 0.0701286569237709, + 0.05580544099211693, + 0.04438965395092964, + 0.10045845061540604, + 0.10688193887472153 + ], + "credit_map": [ + 0.029964277520775795, + 0.02792966365814209, + 0.026633763685822487, + 0.026594283059239388, + 0.019890159368515015, + 0.018405485898256302, + 0.038708966225385666, + 0.025964155793190002, + 0.014390083029866219, + 0.01873672939836979, + 0.015003045089542866, + 0.013948223553597927, + 0.028201716020703316, + 0.04360978677868843, + 0.026579109951853752, + 0.025704212486743927, + 0.00936998799443245, + 0.01146831177175045, + 0.012472300790250301, + 0.013450057245790958, + 0.01423406321555376, + 0.010393224656581879, + 0.013109182007610798, + 0.010626315139234066, + 0.018672997131943703, + 0.01144234836101532, + 0.009884767234325409, + 0.005152763333171606, + 0.0063336752355098724, + 0.009583881124854088, + 0.0064069535583257675, + 0.00845926720649004, + 0.0031414583791047335, + 0.0053239078260958195, + 0.005348267033696175, + 0.00837922003120184, + 0.006541308481246233, + 0.006189980078488588, + 0.009264594875276089, + 0.006823311094194651, + 0.00528733292594552, + 0.014301800169050694, + 0.006661450024694204, + 0.009504028595983982, + 0.00618407828733325, + 0.009422169998288155, + 0.007854696363210678, + 0.010913086123764515, + 0.006339707411825657, + 0.012257291935384274, + 0.006811210419982672, + 0.0062951575964689255, + 0.005540602374821901, + 0.007567915134131908, + 0.006304325070232153, + 0.004689221736043692, + 0.0027864808216691017, + 0.004709760192781687, + 0.0039055771194398403, + 0.0037506953813135624, + 0.007977738045156002, + 0.006694564130157232, + 0.009575361385941505, + 0.004989469423890114, + 0.009399881586432457, + 0.015123491175472736, + 0.00980414915829897, + 0.007727519143372774, + 0.010424784384667873, + 0.016987165436148643, + 0.016764329746365547, + 0.014227115549147129, + 0.007106914650648832, + 0.0078550074249506, + 0.008844367228448391, + 0.008146670646965504, + 0.02223432995378971, + 0.015576724894344807, + 0.018031200394034386, + 0.01908670738339424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.014214574359357357 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282773435115814 + ], + "opposite_logprob_tokens": [ + -2.0178277492523193 + ], + "logprob_reward": -0.14282773435115814, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14282773435115814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.128292977809906, + 0.20008043944835663, + 0.046184465289115906, + 0.1896279901266098, + 0.04209137707948685, + 0.04925408959388733, + 0.04307727515697479, + 0.05853142961859703, + 0.15926863253116608, + 0.08359132707118988 + ], + "credit_map": [ + 0.017769526690244675, + 0.0222404133528471, + 0.014342703856527805, + 0.013076022267341614, + 0.012491986155509949, + 0.013936425559222698, + 0.01891058310866356, + 0.01552532333880663, + 0.01899103820323944, + 0.019171083346009254, + 0.0158171895891428, + 0.013478283770382404, + 0.0324464850127697, + 0.03980076685547829, + 0.030622491613030434, + 0.02975311130285263, + 0.005276121199131012, + 0.0062307268381118774, + 0.005073473323136568, + 0.005922130309045315, + 0.005755000747740269, + 0.005476674996316433, + 0.006882544606924057, + 0.005567794665694237, + 0.04027571156620979, + 0.03284518048167229, + 0.024345604702830315, + 0.016238171607255936, + 0.015619339421391487, + 0.024299904704093933, + 0.014436040073633194, + 0.021568024531006813, + 0.0031502956990152597, + 0.007599495351314545, + 0.0036778999492526054, + 0.0049361116252839565, + 0.0057778796181082726, + 0.005283137783408165, + 0.007083679083734751, + 0.004582878202199936, + 0.004453783854842186, + 0.00865098275244236, + 0.005474607925862074, + 0.007040798664093018, + 0.004605072550475597, + 0.006812578067183495, + 0.005142294801771641, + 0.0070739733055233955, + 0.0043220738880336285, + 0.008034510537981987, + 0.005611653905361891, + 0.005091494414955378, + 0.005031499546021223, + 0.00594853051006794, + 0.004366766195744276, + 0.004670746624469757, + 0.00396574055776, + 0.0055594309233129025, + 0.004919912200421095, + 0.006077615078538656, + 0.010026415809988976, + 0.00822607334703207, + 0.01264942716807127, + 0.007106810342520475, + 0.0118102440610528, + 0.014796651899814606, + 0.013039815239608288, + 0.009947940707206726, + 0.024474723264575005, + 0.0238187238574028, + 0.035310935229063034, + 0.026069600135087967, + 0.004020966589450836, + 0.007715205196291208, + 0.009561493992805481, + 0.007312275003641844, + 0.018900630995631218, + 0.011952027678489685, + 0.012415210716426373, + 0.011713515967130661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.14282773435115814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898349314928055 + ], + "opposite_logprob_tokens": [ + -2.5789835453033447 + ], + "logprob_reward": -0.07898349314928055, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898349314928055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16901931166648865, + 0.08932347595691681, + 0.04649980738759041, + 0.0683414563536644, + 0.02675294689834118, + 0.20773561298847198, + 0.08564164489507675, + 0.04564390704035759, + 0.07560265809297562, + 0.18543924391269684 + ], + "credit_map": [ + 0.022602632641792297, + 0.02357233501970768, + 0.017486579716205597, + 0.02324822172522545, + 0.015423113480210304, + 0.02313910610973835, + 0.01911664381623268, + 0.024430662393569946, + 0.009848288260400295, + 0.014384054578840733, + 0.009209890849888325, + 0.009940536692738533, + 0.011598163284361362, + 0.010994807817041874, + 0.010736539028584957, + 0.012611187063157558, + 0.005445224232971668, + 0.006547224707901478, + 0.004263821057975292, + 0.007258474361151457, + 0.0052155801095068455, + 0.005401906091719866, + 0.006559919100254774, + 0.005807654932141304, + 0.01201226469129324, + 0.011761277914047241, + 0.008210811764001846, + 0.007580388803035021, + 0.0069336239248514175, + 0.008549799211323261, + 0.006313607096672058, + 0.006979681551456451, + 0.002763359108939767, + 0.005550348199903965, + 0.003473552642390132, + 0.003142105881124735, + 0.0027969032526016235, + 0.0037256255745887756, + 0.0027135182172060013, + 0.0025875300634652376, + 0.0208576750010252, + 0.04619864001870155, + 0.01981344446539879, + 0.026316767558455467, + 0.017749521881341934, + 0.030169468373060226, + 0.02191167138516903, + 0.024718403816223145, + 0.010881101712584496, + 0.02178596518933773, + 0.007797114551067352, + 0.010145620442926884, + 0.0068182554095983505, + 0.012686843052506447, + 0.007476243190467358, + 0.008050501346588135, + 0.006116639357060194, + 0.0076586524955928326, + 0.0034652994945645332, + 0.005912054795771837, + 0.004375380463898182, + 0.005958268418908119, + 0.004216833971440792, + 0.007940773852169514, + 0.008465614169836044, + 0.014245256781578064, + 0.008653025142848492, + 0.006978109013289213, + 0.008665534667670727, + 0.010744386352598667, + 0.008615754544734955, + 0.009234965778887272, + 0.020125867798924446, + 0.024675579741597176, + 0.01918661966919899, + 0.014934990555047989, + 0.02447732351720333, + 0.03228391706943512, + 0.024383459240198135, + 0.025371454656124115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.07898349314928055 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062022965401411057 + ], + "opposite_logprob_tokens": [ + -2.8120229244232178 + ], + "logprob_reward": -0.062022965401411057, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062022965401411057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20170274376869202, + 0.06537526845932007, + 0.05438223108649254, + 0.18889452517032623, + 0.04368067905306816, + 0.08589152246713638, + 0.05525661259889603, + 0.08246759325265884, + 0.11584221571683884, + 0.10650663077831268 + ], + "credit_map": [ + 0.024079544469714165, + 0.03175351396203041, + 0.01843542978167534, + 0.025675449520349503, + 0.033326540142297745, + 0.016638392582535744, + 0.02638757787644863, + 0.02540629170835018, + 0.00892865564674139, + 0.006958227138966322, + 0.007637889124453068, + 0.010804335586726665, + 0.006688126362860203, + 0.007040491793304682, + 0.009172780439257622, + 0.008144764229655266, + 0.005902220029383898, + 0.00651684170588851, + 0.005728259216994047, + 0.007525051012635231, + 0.005791507661342621, + 0.006892785429954529, + 0.009655226022005081, + 0.006370340008288622, + 0.046951089054346085, + 0.030491452664136887, + 0.02293754182755947, + 0.012755581177771091, + 0.01623004674911499, + 0.024497108533978462, + 0.015786584466695786, + 0.019245123490691185, + 0.0025512780994176865, + 0.004056381992995739, + 0.005224416498094797, + 0.006629639770835638, + 0.005363771691918373, + 0.0054445709101855755, + 0.008929862640798092, + 0.005480754654854536, + 0.007532327435910702, + 0.010753188282251358, + 0.00871477834880352, + 0.012598306871950626, + 0.008663816377520561, + 0.011878788471221924, + 0.011467017233371735, + 0.01428330596536398, + 0.003708677366375923, + 0.006411789450794458, + 0.005782527849078178, + 0.007260936312377453, + 0.007261359132826328, + 0.007124222815036774, + 0.009470652788877487, + 0.008236448280513287, + 0.007024943828582764, + 0.008028369396924973, + 0.008859810419380665, + 0.00818632822483778, + 0.012211093679070473, + 0.012134331278502941, + 0.015505696646869183, + 0.010517020709812641, + 0.01134649571031332, + 0.017747389152646065, + 0.010921673849225044, + 0.007906503975391388, + 0.015356224030256271, + 0.015542898327112198, + 0.020446980372071266, + 0.01657404564321041, + 0.005047404207289219, + 0.010254355147480965, + 0.011633221991360188, + 0.010594354011118412, + 0.0212965440005064, + 0.017304420471191406, + 0.018601683899760246, + 0.011774643324315548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.062022965401411057 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot depicts the solitary figure of a young man, Dojima, standing motionless on a deserted train platform at night. He has dark, slightly unkempt hair, his face tilted slightly upward, an expression of quiet, melancholic wonder. He wears a simple charcoal-gray hoodie over a T-shirt, his hands tucked deep into the pockets of his dark jeans. The platform itself, made of weathered concrete and empty save for a dark wooden bench, is bathed in the cold, sterile light of a single overhead lamp, which casts long, sharp shadows across the surrounding darkness. Above this silent scene, the universe is in awe-inspiring motion. In a dramatic time-lapse effect, the stars, rather than fixed points, become long, dazzling trails of light, arcing and swirling across the deep indigo canvas of the night sky. The visual contrast between the static, contemplative young man and the frenetic, silent ballet of the universe creates a powerful atmosphere of serene solitude and awe, capturing a single, poignant moment of human existence against the backdrop of cosmic eternity.", + "seed": 7261, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_5/index_6669_seed_7261_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08503273129463196, + "vlm_logprob": -0.08503273129463196, + "vlm_yes_no_margin": 2.574999976158142, + "vlm_token_credit": -0.08503273129463196, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05057620629668236 + ], + "opposite_logprob_tokens": [ + -3.0505762100219727 + ], + "logprob_reward": -0.05057620629668236, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05057620629668236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5105534791946411, + 0.08540285378694534, + 0.049195729196071625, + 0.04027796536684036, + 0.08493203669786453, + 0.09538331627845764, + 0.03235608711838722, + 0.018737390637397766, + 0.032359179109334946, + 0.05080195143818855 + ], + "credit_map": [ + 0.05363171547651291, + 0.054523978382349014, + 0.07020765542984009, + 0.053820621222257614, + 0.07682513445615768, + 0.07714200764894485, + 0.05770367383956909, + 0.06669871509075165, + 0.007727127522230148, + 0.006056578829884529, + 0.01575995795428753, + 0.011898238211870193, + 0.006789103150367737, + 0.01947629451751709, + 0.008846393786370754, + 0.008849170990288258, + 0.0043076323345303535, + 0.005064808763563633, + 0.008917292580008507, + 0.005199640057981014, + 0.004562947899103165, + 0.009974317625164986, + 0.006310644559562206, + 0.004858449567109346, + 0.006348780822008848, + 0.004036158323287964, + 0.006514813285320997, + 0.0037988037802278996, + 0.0035760030150413513, + 0.0050187064334750175, + 0.004915149882435799, + 0.00606955261901021, + 0.010057775303721428, + 0.007762204855680466, + 0.010101945139467716, + 0.012442749924957752, + 0.007860159501433372, + 0.010336305014789104, + 0.011630036868155003, + 0.014740870334208012, + 0.012014977633953094, + 0.01478538941591978, + 0.01580423302948475, + 0.010480375029146671, + 0.006792089901864529, + 0.01694420352578163, + 0.009029019623994827, + 0.009533037431538105, + 0.001975002931430936, + 0.0032446770928800106, + 0.0040077343583106995, + 0.0037899406161159277, + 0.004078215453773737, + 0.0040273540653288364, + 0.007136098574846983, + 0.004097066354006529, + 0.002204534597694874, + 0.0017491207690909505, + 0.002181870164349675, + 0.002070131478831172, + 0.0021740489173680544, + 0.0022142529487609863, + 0.003928931895643473, + 0.0022145011462271214, + 0.004197454545646906, + 0.004412262234836817, + 0.0039122034795582294, + 0.003921739291399717, + 0.004528346937149763, + 0.0038993647322058678, + 0.004158304538577795, + 0.0033295087050646544, + 0.008663010783493519, + 0.0073258960619568825, + 0.005586362909525633, + 0.00457562692463398, + 0.005334174260497093, + 0.006634221877902746, + 0.006275692488998175, + 0.006406969856470823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.05057620629668236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981521561741829 + ], + "opposite_logprob_tokens": [ + -3.529815196990967 + ], + "logprob_reward": -0.02981521561741829, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981521561741829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17768891155719757, + 0.10218409448862076, + 0.07302065193653107, + 0.045318443328142166, + 0.06392516195774078, + 0.12441523373126984, + 0.0718001201748848, + 0.15364794433116913, + 0.03379623964428902, + 0.15420323610305786 + ], + "credit_map": [ + 0.015934854745864868, + 0.016377417370676994, + 0.02124081365764141, + 0.020905274897813797, + 0.02766394056379795, + 0.02436116710305214, + 0.024455569684505463, + 0.02674986980855465, + 0.010410800576210022, + 0.008647242560982704, + 0.020416870713233948, + 0.00796019658446312, + 0.013091661036014557, + 0.025691207498311996, + 0.007318640127778053, + 0.00864747166633606, + 0.00601238664239645, + 0.006542694754898548, + 0.01559448428452015, + 0.007562029641121626, + 0.007453161291778088, + 0.013825559988617897, + 0.008275732398033142, + 0.007754595018923283, + 0.005207844078540802, + 0.004712969530373812, + 0.0062654088251292706, + 0.003836557501927018, + 0.006724837701767683, + 0.005044619552791119, + 0.005874049849808216, + 0.007652156986296177, + 0.00757029140368104, + 0.006177747622132301, + 0.006874947343021631, + 0.006804842036217451, + 0.008379924111068249, + 0.006922850385308266, + 0.010083665139973164, + 0.011110898107290268, + 0.010692364536225796, + 0.016504552215337753, + 0.01902342028915882, + 0.014219900593161583, + 0.01479795016348362, + 0.02535047009587288, + 0.011578859761357307, + 0.01224772073328495, + 0.004916519392281771, + 0.0054331207647919655, + 0.007641250733286142, + 0.007806098088622093, + 0.013808720745146275, + 0.008340497501194477, + 0.01505217794328928, + 0.008801735006272793, + 0.02089490182697773, + 0.01340918242931366, + 0.020530540496110916, + 0.013512251898646355, + 0.01629919745028019, + 0.0211341455578804, + 0.02486647479236126, + 0.02300124615430832, + 0.0041416543535888195, + 0.004893102217465639, + 0.004739804193377495, + 0.0037901438772678375, + 0.0034142020158469677, + 0.005200101528316736, + 0.004173682536929846, + 0.0034435491543263197, + 0.00991267990320921, + 0.01740175485610962, + 0.01966746710240841, + 0.014293069951236248, + 0.024405188858509064, + 0.01907493732869625, + 0.028605392202734947, + 0.02084273472428322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.02981521561741829 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294959485530853 + ], + "opposite_logprob_tokens": [ + -2.2379496097564697 + ], + "logprob_reward": -0.11294959485530853, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294959485530853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1400064378976822, + 0.07022526115179062, + 0.06442588567733765, + 0.10169607400894165, + 0.03334122523665428, + 0.26170697808265686, + 0.04535563662648201, + 0.04689319059252739, + 0.02934778854250908, + 0.2070014923810959 + ], + "credit_map": [ + 0.011968851089477539, + 0.014058778062462807, + 0.01790132001042366, + 0.016544831916689873, + 0.027930909767746925, + 0.014184873551130295, + 0.02205742336809635, + 0.015359454788267612, + 0.007643697317689657, + 0.005307035520672798, + 0.012691667303442955, + 0.005700850859284401, + 0.011183238588273525, + 0.014261377044022083, + 0.00758423050865531, + 0.005853166338056326, + 0.008584477938711643, + 0.006220662500709295, + 0.008017833344638348, + 0.005895692389458418, + 0.007512382231652737, + 0.010122657753527164, + 0.009959239512681961, + 0.00811294186860323, + 0.018231483176350594, + 0.013489742763340473, + 0.012587714940309525, + 0.0076914140954613686, + 0.010787563398480415, + 0.015015381388366222, + 0.009561076760292053, + 0.014331698417663574, + 0.0033321904484182596, + 0.003507157089188695, + 0.004463139921426773, + 0.003958546556532383, + 0.00465658213943243, + 0.0038531324826180935, + 0.005442644935101271, + 0.004127833526581526, + 0.03192351013422012, + 0.049956824630498886, + 0.045361194759607315, + 0.024506330490112305, + 0.021305466070771217, + 0.030935006216168404, + 0.0249821487814188, + 0.03273649886250496, + 0.003559666918590665, + 0.0040480466559529305, + 0.005854160990566015, + 0.0049985055811703205, + 0.0065298425033688545, + 0.00657768826931715, + 0.00755548058077693, + 0.006232244428247213, + 0.006348087918013334, + 0.004179850686341524, + 0.005408451426774263, + 0.0046723405830562115, + 0.0072735571302473545, + 0.006715778727084398, + 0.007690697442740202, + 0.004604429006576538, + 0.0033426375593990088, + 0.004186118021607399, + 0.0035860182251781225, + 0.0027370061725378036, + 0.003879191353917122, + 0.0034739819820970297, + 0.004802759736776352, + 0.0033400761894881725, + 0.04481144994497299, + 0.024726077914237976, + 0.029246507212519646, + 0.015474173240363598, + 0.016563311219215393, + 0.02693675458431244, + 0.023337814956903458, + 0.025905396789312363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.11294959485530853 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904317021369934 + ], + "opposite_logprob_tokens": [ + -2.464043140411377 + ], + "logprob_reward": -0.08904317021369934, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904317021369934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5452216267585754, + 0.06508437544107437, + 0.04659949988126755, + 0.04007332772016525, + 0.0634138435125351, + 0.14435555040836334, + 0.012819476425647736, + 0.018967049196362495, + 0.014420541003346443, + 0.049044691026210785 + ], + "credit_map": [ + 0.059584807604551315, + 0.05382025986909866, + 0.0721956342458725, + 0.05713969096541405, + 0.08839458972215652, + 0.08282005786895752, + 0.06590650230646133, + 0.06536011397838593, + 0.007231270428746939, + 0.004611191805452108, + 0.013734645210206509, + 0.005327880848199129, + 0.006117923650890589, + 0.016840357333421707, + 0.004817086737602949, + 0.006404017563909292, + 0.00579818245023489, + 0.006529327481985092, + 0.008087498135864735, + 0.0045688338577747345, + 0.0037691800389438868, + 0.00986509583890438, + 0.0037221580278128386, + 0.004259222652763128, + 0.004954722709953785, + 0.004707405809313059, + 0.0057503823190927505, + 0.004093989729881287, + 0.0042669582180678844, + 0.006681214552372694, + 0.004235329572111368, + 0.005383328069001436, + 0.006794476881623268, + 0.006193735171109438, + 0.006951033603399992, + 0.008384919725358486, + 0.006064092740416527, + 0.008408617228269577, + 0.009361105039715767, + 0.011255857534706593, + 0.01873120106756687, + 0.022370733320713043, + 0.022189825773239136, + 0.016669271513819695, + 0.010546568781137466, + 0.02494250424206257, + 0.013158204965293407, + 0.01574724167585373, + 0.0011159833520650864, + 0.001428551273420453, + 0.001965523697435856, + 0.0012997243320569396, + 0.0019229220924898982, + 0.0019850023090839386, + 0.0016845533391460776, + 0.0014172163791954517, + 0.0015920872101560235, + 0.0016560512594878674, + 0.0023036147467792034, + 0.002074087504297495, + 0.0029057522770017385, + 0.003288879757747054, + 0.0027435992378741503, + 0.002402977552264929, + 0.0015737422509118915, + 0.0016738924896344543, + 0.002001581247895956, + 0.0010067932307720184, + 0.001888797152787447, + 0.002917557954788208, + 0.0014845788246020675, + 0.0018735980847850442, + 0.010248299688100815, + 0.005901484750211239, + 0.007690176367759705, + 0.0034312698990106583, + 0.0028492463752627373, + 0.009552503935992718, + 0.004786951933056116, + 0.004584759008139372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.08904317021369934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277946949005127 + ], + "opposite_logprob_tokens": [ + -2.0177793502807617 + ], + "logprob_reward": -0.14277946949005127, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14277946949005127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27941033244132996, + 0.057409852743148804, + 0.04574333131313324, + 0.08049071580171585, + 0.05509723350405693, + 0.23085913062095642, + 0.0582311749458313, + 0.030302666127681732, + 0.04115724191069603, + 0.12129829078912735 + ], + "credit_map": [ + 0.02437729388475418, + 0.02975899539887905, + 0.034172721207141876, + 0.033166058361530304, + 0.047973036766052246, + 0.03699677065014839, + 0.03795583173632622, + 0.03500962629914284, + 0.009019087068736553, + 0.004710554610937834, + 0.010832657106220722, + 0.003713924903422594, + 0.00533456401899457, + 0.014348512515425682, + 0.004548126831650734, + 0.004902427550405264, + 0.0060841781087219715, + 0.005129263736307621, + 0.007095668464899063, + 0.004047316499054432, + 0.004931827075779438, + 0.009025591425597668, + 0.004613551776856184, + 0.004815932363271713, + 0.01449598092585802, + 0.010226037353277206, + 0.010846599005162716, + 0.006253010127693415, + 0.008904744870960712, + 0.011772857047617435, + 0.007481918204575777, + 0.010509572923183441, + 0.007479448802769184, + 0.0052644615061581135, + 0.008282430469989777, + 0.005343882367014885, + 0.0069143520668148994, + 0.005357459187507629, + 0.006845502648502588, + 0.009609694592654705, + 0.027253882959485054, + 0.043981462717056274, + 0.03796776756644249, + 0.021570241078734398, + 0.018786080181598663, + 0.030735932290554047, + 0.019552936777472496, + 0.031010804697871208, + 0.00465997401624918, + 0.004817119333893061, + 0.0077518438920378685, + 0.006230250000953674, + 0.007731281686574221, + 0.009174559265375137, + 0.009211406111717224, + 0.008654735051095486, + 0.003705644514411688, + 0.0032033149618655443, + 0.0029697581194341183, + 0.0024208356626331806, + 0.004382086452096701, + 0.0037739952094852924, + 0.005741605535149574, + 0.004105426371097565, + 0.004211848601698875, + 0.00682791555300355, + 0.00455750385299325, + 0.0038113845512270927, + 0.00574845913797617, + 0.005337520036846399, + 0.0064491513185203075, + 0.004213459324091673, + 0.030460581183433533, + 0.01495566125959158, + 0.016400517895817757, + 0.008624309673905373, + 0.009868493303656578, + 0.015819206833839417, + 0.011576120741665363, + 0.013593406416475773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.14277946949005127 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot depicts the solitary figure of a young man, Dojima, standing motionless on a deserted train platform at night. He has dark, slightly unkempt hair, his face tilted slightly upward, an expression of quiet, melancholic wonder. He wears a simple charcoal-gray hoodie over a T-shirt, his hands tucked deep into the pockets of his dark jeans. The platform itself, made of weathered concrete and empty save for a dark wooden bench, is bathed in the cold, sterile light of a single overhead lamp, which casts long, sharp shadows across the surrounding darkness. Above this silent scene, the universe is in awe-inspiring motion. In a dramatic time-lapse effect, the stars, rather than fixed points, become long, dazzling trails of light, arcing and swirling across the deep indigo canvas of the night sky. The visual contrast between the static, contemplative young man and the frenetic, silent ballet of the universe creates a powerful atmosphere of serene solitude and awe, capturing a single, poignant moment of human existence against the backdrop of cosmic eternity.", + "seed": 7262, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_5/index_6669_seed_7262_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.022845768742263316, + "vlm_logprob": -0.022845768742263316, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.022845768742263316, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013894615694880486 + ], + "opposite_logprob_tokens": [ + -4.513894557952881 + ], + "logprob_reward": -0.013894615694880486, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013894615694880486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18633008003234863, + 0.23802387714385986, + 0.18117433786392212, + 0.13328424096107483, + 0.05967509374022484, + 0.02491355687379837, + 0.029676096513867378, + 0.04872029274702072, + 0.059074290096759796, + 0.03912810981273651 + ], + "credit_map": [ + 0.02725696563720703, + 0.028817452490329742, + 0.039588212966918945, + 0.01705607771873474, + 0.01391674391925335, + 0.026253756135702133, + 0.013584339059889317, + 0.019856523722410202, + 0.04509275406599045, + 0.09437660872936249, + 0.015069701708853245, + 0.010689790360629559, + 0.011818730272352695, + 0.04054311662912369, + 0.007568364962935448, + 0.012864815071225166, + 0.01953951083123684, + 0.040534596890211105, + 0.027974486351013184, + 0.023918401449918747, + 0.014038730412721634, + 0.022993069142103195, + 0.01462707668542862, + 0.017548469826579094, + 0.017199700698256493, + 0.02558194287121296, + 0.019065456464886665, + 0.01348638441413641, + 0.012092243880033493, + 0.02229895070195198, + 0.010299112647771835, + 0.01326045859605074, + 0.007452286314219236, + 0.012170995585620403, + 0.005910723004490137, + 0.007256242446601391, + 0.0063933804631233215, + 0.0073516094125807285, + 0.007265746593475342, + 0.005874113645404577, + 0.0035171308554708958, + 0.004299847409129143, + 0.00369037757627666, + 0.0026359395124018192, + 0.002274325815960765, + 0.0029308716766536236, + 0.0026212921366095543, + 0.0029437714256346226, + 0.0037745737936347723, + 0.004587571136653423, + 0.0048861755058169365, + 0.0038201832212507725, + 0.0023768094833940268, + 0.003610232612118125, + 0.002780672162771225, + 0.0038398790638893843, + 0.005394331179559231, + 0.010070445947349072, + 0.006735502742230892, + 0.004986763931810856, + 0.0036192249972373247, + 0.006650266237556934, + 0.003753542434424162, + 0.0075102136470377445, + 0.007324695587158203, + 0.009094752371311188, + 0.00799327902495861, + 0.0073680272325873375, + 0.00479477783665061, + 0.008378760889172554, + 0.004808780271559954, + 0.009311215952038765, + 0.005255137104541063, + 0.008081840351223946, + 0.005489610601216555, + 0.004078634548932314, + 0.0030872421339154243, + 0.006352418102324009, + 0.0033707302063703537, + 0.0034124976955354214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.013894615694880486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297740146517754 + ], + "opposite_logprob_tokens": [ + -3.7732977867126465 + ], + "logprob_reward": -0.023297740146517754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297740146517754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0917418971657753, + 0.6356735229492188, + 0.033321283757686615, + 0.03046184591948986, + 0.08368992805480957, + 0.02611652761697769, + 0.0222417414188385, + 0.01610834151506424, + 0.014505569823086262, + 0.0461394339799881 + ], + "credit_map": [ + 0.010336693376302719, + 0.013480434194207191, + 0.0091493995860219, + 0.006911293603479862, + 0.010801723226904869, + 0.024932051077485085, + 0.007624083664268255, + 0.00850620772689581, + 0.12620852887630463, + 0.26557475328445435, + 0.02719617262482643, + 0.026281118392944336, + 0.02596343494951725, + 0.11542143672704697, + 0.017124302685260773, + 0.031903691589832306, + 0.0032093296758830547, + 0.0052713858895003796, + 0.0020862091332674026, + 0.0034135247115045786, + 0.006191803142428398, + 0.0037293797358870506, + 0.005005122162401676, + 0.004414524883031845, + 0.004864857066422701, + 0.003802158636972308, + 0.0032141644041985273, + 0.002069479087367654, + 0.004834991414099932, + 0.006697517819702625, + 0.002928120084106922, + 0.0020505532156676054, + 0.009056990034878254, + 0.008540407754480839, + 0.008803709410130978, + 0.005954091902822256, + 0.020663924515247345, + 0.010305233299732208, + 0.01404547318816185, + 0.006320087239146233, + 0.002851728582754731, + 0.004371405113488436, + 0.002376994350925088, + 0.002032183576375246, + 0.005570875480771065, + 0.002814893377944827, + 0.003957692999392748, + 0.00214075087569654, + 0.0023576414678245783, + 0.003252929775044322, + 0.0027934552635997534, + 0.002539839828386903, + 0.003536748234182596, + 0.002732846885919571, + 0.0025321815628558397, + 0.002496096072718501, + 0.0016031452687457204, + 0.0025041389744728804, + 0.0018071227241307497, + 0.0021931955125182867, + 0.0016276363749057055, + 0.0021305636037141085, + 0.0020030306186527014, + 0.0022395062260329723, + 0.0019230942707508802, + 0.003277532057836652, + 0.0021986004430800676, + 0.001136883394792676, + 0.0014258814044296741, + 0.0019783673342317343, + 0.001057733898051083, + 0.0015074757393449545, + 0.006117579992860556, + 0.006733256857842207, + 0.006029838230460882, + 0.0032668844796717167, + 0.0061713457107543945, + 0.007109238300472498, + 0.0056228479370474815, + 0.005088437348604202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.023297740146517754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033705130219459534 + ], + "opposite_logprob_tokens": [ + -3.408705234527588 + ], + "logprob_reward": -0.033705130219459534, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033705130219459534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11245110630989075, + 0.678521454334259, + 0.027142783626914024, + 0.03143090382218361, + 0.019369814544916153, + 0.04386791214346886, + 0.013853183947503567, + 0.029255123808979988, + 0.023620987311005592, + 0.02048676274716854 + ], + "credit_map": [ + 0.021902484819293022, + 0.011004920117557049, + 0.02109501138329506, + 0.009506434202194214, + 0.011473115533590317, + 0.014867465943098068, + 0.010513697750866413, + 0.012087972834706306, + 0.13518112897872925, + 0.2847414016723633, + 0.029847223311662674, + 0.028244050219655037, + 0.027210121974349022, + 0.121686652302742, + 0.01771599054336548, + 0.033894915133714676, + 0.0031154982279986143, + 0.005878262221813202, + 0.0030400711111724377, + 0.003086106153205037, + 0.003732989076524973, + 0.0032291505485773087, + 0.0023298701271414757, + 0.0027308366261422634, + 0.0036607945803552866, + 0.005441396497189999, + 0.0032155937515199184, + 0.002942649880424142, + 0.005401591304689646, + 0.004852999933063984, + 0.003540611593052745, + 0.0023752632550895214, + 0.0025748549960553646, + 0.004022086970508099, + 0.0015812978381291032, + 0.0026815372984856367, + 0.0022552735172212124, + 0.0021764500997960567, + 0.0022392987739294767, + 0.001839015050791204, + 0.005976267624646425, + 0.013373197056353092, + 0.0036848424933850765, + 0.003775383345782757, + 0.004024488851428032, + 0.005816471762955189, + 0.0029987806919962168, + 0.004218478687107563, + 0.0012334546772763133, + 0.002144297119230032, + 0.0015246154507622123, + 0.001527042593806982, + 0.0021053224336355925, + 0.001914618769660592, + 0.001799122546799481, + 0.001604710123501718, + 0.003718065330758691, + 0.006555439438670874, + 0.002683510072529316, + 0.0022391830570995808, + 0.00302337110042572, + 0.004127270542085171, + 0.0024896434042602777, + 0.004418640397489071, + 0.002724683377891779, + 0.003902577795088291, + 0.0018585947109386325, + 0.0017467609141021967, + 0.0034087488893419504, + 0.004004235845059156, + 0.003023067256435752, + 0.002952317940071225, + 0.002781316637992859, + 0.0036533104721456766, + 0.002722651232033968, + 0.0017085425788536668, + 0.0024482363369315863, + 0.003058219561353326, + 0.0021353312768042088, + 0.001979153836145997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.033705130219459534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250236950814724 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.005250236950814724, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250236950814724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11234483122825623, + 0.15804941952228546, + 0.08997642248868942, + 0.07989856600761414, + 0.08167243003845215, + 0.06555825471878052, + 0.055961694568395615, + 0.2654806673526764, + 0.030576979741454124, + 0.0604807510972023 + ], + "credit_map": [ + 0.018761593848466873, + 0.015320851467549801, + 0.012142129242420197, + 0.012956146150827408, + 0.008595459163188934, + 0.019086571410298347, + 0.010253620333969593, + 0.0152284512296319, + 0.028307700529694557, + 0.05732506886124611, + 0.012723159044981003, + 0.010204252786934376, + 0.008743170648813248, + 0.0257481187582016, + 0.006484746001660824, + 0.008513197302818298, + 0.010606438852846622, + 0.01895112358033657, + 0.012365966103971004, + 0.013513616286218166, + 0.007236852776259184, + 0.011723227798938751, + 0.008205961436033249, + 0.007373237516731024, + 0.014071769081056118, + 0.011119970120489597, + 0.008691503666341305, + 0.006828072015196085, + 0.006093032658100128, + 0.01910245232284069, + 0.006629284005612135, + 0.007362483534961939, + 0.010364688001573086, + 0.019480500370264053, + 0.0058780754916369915, + 0.01008539367467165, + 0.008327544666826725, + 0.010907364077866077, + 0.009732984937727451, + 0.006895882543176413, + 0.008733833208680153, + 0.013849154114723206, + 0.007619356736540794, + 0.008798187598586082, + 0.005928222090005875, + 0.0077522629871964455, + 0.005573441740125418, + 0.007303793914616108, + 0.006417552474886179, + 0.009224114939570427, + 0.009774425067007542, + 0.006287586409598589, + 0.004516012035310268, + 0.007968018762767315, + 0.00538982916623354, + 0.006384157110005617, + 0.028950169682502747, + 0.07013171166181564, + 0.02655828557908535, + 0.024358408525586128, + 0.019615070894360542, + 0.044811420142650604, + 0.020583229139447212, + 0.030472347512841225, + 0.003954026848077774, + 0.006798872258514166, + 0.005090855527669191, + 0.004153126385062933, + 0.0016296447720378637, + 0.004380444996058941, + 0.0021227996330708265, + 0.0024472095537930727, + 0.007345371413975954, + 0.01479465514421463, + 0.007484509609639645, + 0.005444481037557125, + 0.004249835852533579, + 0.010624694637954235, + 0.005283779930323362, + 0.00525342533364892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.005250236950814724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808112069964409 + ], + "opposite_logprob_tokens": [ + -3.288081169128418 + ], + "logprob_reward": -0.03808112069964409, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808112069964409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21950741112232208, + 0.08439620584249496, + 0.10805345326662064, + 0.26004645228385925, + 0.06303417682647705, + 0.07275521755218506, + 0.04693803936243057, + 0.05922936648130417, + 0.04385850206017494, + 0.04218119755387306 + ], + "credit_map": [ + 0.030360814183950424, + 0.032207440584897995, + 0.028019573539495468, + 0.020529640838503838, + 0.024659819900989532, + 0.04729441553354263, + 0.0201212577521801, + 0.016314459964632988, + 0.01088879257440567, + 0.016228679567575455, + 0.009865453466773033, + 0.007208008784800768, + 0.011415966786444187, + 0.013244999572634697, + 0.0077257947996258736, + 0.00781851727515459, + 0.013018917292356491, + 0.020350245758891106, + 0.015237056650221348, + 0.014539970085024834, + 0.011381582356989384, + 0.015267407521605492, + 0.010374150238931179, + 0.007884127087891102, + 0.06517717987298965, + 0.026378683745861053, + 0.012585964985191822, + 0.010225887410342693, + 0.015623465180397034, + 0.0924048125743866, + 0.014952789060771465, + 0.022697657346725464, + 0.007039979565888643, + 0.006936612073332071, + 0.007608623709529638, + 0.007205905392765999, + 0.012226582504808903, + 0.007594830822199583, + 0.0094815818592906, + 0.0049400669522583485, + 0.008044702932238579, + 0.016305962577462196, + 0.008481401950120926, + 0.0074207475408911705, + 0.008444810286164284, + 0.008464813232421875, + 0.006508619524538517, + 0.009084157645702362, + 0.006143881008028984, + 0.006534151267260313, + 0.004506452474743128, + 0.00556968804448843, + 0.008572064340114594, + 0.005807200912386179, + 0.0061630974523723125, + 0.003641504095867276, + 0.005882697645574808, + 0.009207328781485558, + 0.006289216689765453, + 0.008631641045212746, + 0.007667504716664553, + 0.008565378375351429, + 0.0060492646880447865, + 0.006936336867511272, + 0.0064277565106749535, + 0.007063424680382013, + 0.0047620609402656555, + 0.006679905112832785, + 0.004588007926940918, + 0.005688718520104885, + 0.004427869338542223, + 0.00422075716778636, + 0.006304461974650621, + 0.00928270909935236, + 0.004188039805740118, + 0.0039446656592190266, + 0.003830968402326107, + 0.0073655168525874615, + 0.003277785377576947, + 0.003987051080912352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.03808112069964409 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot depicts the solitary figure of a young man, Dojima, standing motionless on a deserted train platform at night. He has dark, slightly unkempt hair, his face tilted slightly upward, an expression of quiet, melancholic wonder. He wears a simple charcoal-gray hoodie over a T-shirt, his hands tucked deep into the pockets of his dark jeans. The platform itself, made of weathered concrete and empty save for a dark wooden bench, is bathed in the cold, sterile light of a single overhead lamp, which casts long, sharp shadows across the surrounding darkness. Above this silent scene, the universe is in awe-inspiring motion. In a dramatic time-lapse effect, the stars, rather than fixed points, become long, dazzling trails of light, arcing and swirling across the deep indigo canvas of the night sky. The visual contrast between the static, contemplative young man and the frenetic, silent ballet of the universe creates a powerful atmosphere of serene solitude and awe, capturing a single, poignant moment of human existence against the backdrop of cosmic eternity.", + "seed": 7263, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_5/index_6669_seed_7263_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06442482471466064, + "vlm_logprob": -0.06442482471466064, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.06442482471466064, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045233067125082016 + ], + "opposite_logprob_tokens": [ + -3.1702330112457275 + ], + "logprob_reward": -0.045233067125082016, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.045233067125082016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4096412658691406, + 0.07608961313962936, + 0.014505444094538689, + 0.1063474714756012, + 0.0479825995862484, + 0.0774858295917511, + 0.05133603513240814, + 0.12520001828670502, + 0.027416285127401352, + 0.06399541348218918 + ], + "credit_map": [ + 0.060505833476781845, + 0.058386482298374176, + 0.05281712859869003, + 0.031229764223098755, + 0.059756841510534286, + 0.06077421456575394, + 0.05917994678020477, + 0.026991089805960655, + 0.011015613563358784, + 0.009726336225867271, + 0.012059150263667107, + 0.006503722630441189, + 0.007006423082202673, + 0.015296382829546928, + 0.0060006845742464066, + 0.008481298573315144, + 0.0018050340004265308, + 0.0015441358555108309, + 0.0018495189724490047, + 0.0014952089404687285, + 0.0017175426473841071, + 0.002296139718964696, + 0.0018423267174512148, + 0.001955537125468254, + 0.027947047725319862, + 0.014712599106132984, + 0.015024554915726185, + 0.011608630418777466, + 0.007370830047875643, + 0.015143203549087048, + 0.0075679984875023365, + 0.00697260070592165, + 0.008389757946133614, + 0.008658185601234436, + 0.007592897396534681, + 0.005475286394357681, + 0.004112078342586756, + 0.0068361880257725716, + 0.003941421862691641, + 0.002976781688630581, + 0.015857547521591187, + 0.020123109221458435, + 0.008830361068248749, + 0.006476166658103466, + 0.005033738445490599, + 0.011865077540278435, + 0.004554727580398321, + 0.00474510295316577, + 0.009908092208206654, + 0.013091032393276691, + 0.0048643252812325954, + 0.004787703510373831, + 0.003160822903737426, + 0.0073037538677453995, + 0.0037506455555558205, + 0.00446966290473938, + 0.01896827667951584, + 0.022499240934848785, + 0.015165559016168118, + 0.018333425745368004, + 0.009440123103559017, + 0.016012288630008698, + 0.011532850563526154, + 0.013248251751065254, + 0.004502372816205025, + 0.002914732089266181, + 0.0031599970534443855, + 0.002149182138964534, + 0.004372583236545324, + 0.003426539245992899, + 0.004450170323252678, + 0.0024407084565609694, + 0.009851610288023949, + 0.009374643675982952, + 0.008445486426353455, + 0.007747781928628683, + 0.005208227783441544, + 0.008023749105632305, + 0.006917251273989677, + 0.008426662534475327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.045233067125082016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059445157647133 + ], + "opposite_logprob_tokens": [ + -3.895594358444214 + ], + "logprob_reward": -0.02059445157647133, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059445157647133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2650470435619354, + 0.07304047793149948, + 0.06883318722248077, + 0.09129384905099869, + 0.0752558559179306, + 0.05825979262590408, + 0.11220663785934448, + 0.1484188735485077, + 0.0632471963763237, + 0.04439707472920418 + ], + "credit_map": [ + 0.028013678267598152, + 0.02612646482884884, + 0.02665247954428196, + 0.027787741273641586, + 0.052165545523166656, + 0.036488667130470276, + 0.04681987315416336, + 0.020992599427700043, + 0.008177234791219234, + 0.010486294515430927, + 0.008031419478356838, + 0.0058822291903197765, + 0.013997041620314121, + 0.008334220387041569, + 0.012629331089556217, + 0.005502712447196245, + 0.007343758828938007, + 0.006884211674332619, + 0.006561602000147104, + 0.00741637172177434, + 0.014188348315656185, + 0.006331665441393852, + 0.014972522854804993, + 0.005134705454111099, + 0.013104522600769997, + 0.016693411394953728, + 0.008724426850676537, + 0.006483874283730984, + 0.01081517431885004, + 0.015061723068356514, + 0.013197442516684532, + 0.007213280536234379, + 0.008539975620806217, + 0.008579603396356106, + 0.010189982131123543, + 0.00882736686617136, + 0.011236073449254036, + 0.007421422749757767, + 0.013330148532986641, + 0.007131289690732956, + 0.008922269567847252, + 0.013952000066637993, + 0.006352387834340334, + 0.005521685816347599, + 0.0055196089670062065, + 0.007550783921033144, + 0.005131827667355537, + 0.005309229716658592, + 0.0211939737200737, + 0.0223408080637455, + 0.016115112230181694, + 0.010852665640413761, + 0.009103290736675262, + 0.015530789270997047, + 0.008148660883307457, + 0.00892132893204689, + 0.02457662671804428, + 0.024504249915480614, + 0.01898186281323433, + 0.022337621077895164, + 0.011725305579602718, + 0.018450025469064713, + 0.014466273598372936, + 0.013376905582845211, + 0.008810439147055149, + 0.008002230897545815, + 0.009451792575418949, + 0.005225450731813908, + 0.00919373333454132, + 0.007352438755333424, + 0.009090587496757507, + 0.00612052483484149, + 0.00573712307959795, + 0.005339445546269417, + 0.006020238157361746, + 0.004725800361484289, + 0.00471093226224184, + 0.004755248315632343, + 0.007522460538893938, + 0.005585823208093643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.02059445157647133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032714158296585 + ], + "opposite_logprob_tokens": [ + -2.350327253341675 + ], + "logprob_reward": -0.10032714158296585, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032714158296585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31856977939605713, + 0.05076777935028076, + 0.037229783833026886, + 0.07447336614131927, + 0.06738230586051941, + 0.055355221033096313, + 0.189701646566391, + 0.12097945064306259, + 0.044286277145147324, + 0.041254375129938126 + ], + "credit_map": [ + 0.057304296642541885, + 0.0459773875772953, + 0.05047185346484184, + 0.031047463417053223, + 0.03847067058086395, + 0.03580395132303238, + 0.04053780063986778, + 0.018956368789076805, + 0.0051260231994092464, + 0.006244537886232138, + 0.006407120730727911, + 0.005170821212232113, + 0.007370851468294859, + 0.009192506782710552, + 0.006550519727170467, + 0.00470540439710021, + 0.0038563937414437532, + 0.004315007943660021, + 0.004224951844662428, + 0.003807738656178117, + 0.0059441193006932735, + 0.004268095828592777, + 0.0069211432710289955, + 0.0038923376705497503, + 0.009001593105494976, + 0.010287057608366013, + 0.006511623039841652, + 0.005119995214045048, + 0.013131452724337578, + 0.014566374942660332, + 0.011022648774087429, + 0.004832630045711994, + 0.010216502472758293, + 0.010053064674139023, + 0.009458648972213268, + 0.006262696348130703, + 0.009091222658753395, + 0.007792365271598101, + 0.010389195755124092, + 0.004118614364415407, + 0.0109608955681324, + 0.01389478612691164, + 0.005673266481608152, + 0.004641377367079258, + 0.00567604461684823, + 0.007492233999073505, + 0.0042636520229279995, + 0.0027529720682650805, + 0.030967511236667633, + 0.024439044296741486, + 0.02706972509622574, + 0.01603424735367298, + 0.029060836881399155, + 0.019598141312599182, + 0.02791816182434559, + 0.0146139832213521, + 0.012541178613901138, + 0.018566148355603218, + 0.012531266547739506, + 0.01905018277466297, + 0.014922854490578175, + 0.013371528126299381, + 0.017707619816064835, + 0.012288682162761688, + 0.007330269552767277, + 0.00459652952849865, + 0.004924453794956207, + 0.0036550890654325485, + 0.006893491838127375, + 0.005190791096538305, + 0.006939208135008812, + 0.004756446927785873, + 0.005650966428220272, + 0.004693097900599241, + 0.005556038580834866, + 0.005136936902999878, + 0.005119833163917065, + 0.004430434666574001, + 0.0050880033522844315, + 0.005579069722443819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.10032714158296585 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306214302778244 + ], + "opposite_logprob_tokens": [ + -3.168062210083008 + ], + "logprob_reward": -0.04306214302778244, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306214302778244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49983489513397217, + 0.04261339083313942, + 0.016941748559474945, + 0.05112416669726372, + 0.060710664838552475, + 0.035235319286584854, + 0.08105005323886871, + 0.12423171103000641, + 0.03486214578151703, + 0.053395893424749374 + ], + "credit_map": [ + 0.07767773419618607, + 0.07577291131019592, + 0.0756908655166626, + 0.03868206590414047, + 0.06755046546459198, + 0.07260526716709137, + 0.06106828898191452, + 0.030787335708737373, + 0.00571496831253171, + 0.005846515763550997, + 0.006738147232681513, + 0.005043366923928261, + 0.0044453339651227, + 0.008343344554305077, + 0.003275043796747923, + 0.0032066688872873783, + 0.0022063523065298796, + 0.0017222721362486482, + 0.0030151440296322107, + 0.00242016464471817, + 0.0015760912792757154, + 0.00213254033587873, + 0.0023445056285709143, + 0.001524677500128746, + 0.012647271156311035, + 0.007785791531205177, + 0.005714389029890299, + 0.004881883505731821, + 0.004221459850668907, + 0.006386917550116777, + 0.005424422677606344, + 0.004062030930072069, + 0.01255941204726696, + 0.010997897945344448, + 0.009515675716102123, + 0.006573671940714121, + 0.0036353757604956627, + 0.008680573664605618, + 0.004443808924406767, + 0.004304247908294201, + 0.007646228652447462, + 0.00805683434009552, + 0.004202771000564098, + 0.0026228846982121468, + 0.002338915830478072, + 0.004895979072898626, + 0.002247849479317665, + 0.003223853651434183, + 0.01765792816877365, + 0.016682378947734833, + 0.011539322324097157, + 0.00873875617980957, + 0.004927973262965679, + 0.011133825406432152, + 0.004546106792986393, + 0.005823764484375715, + 0.02079910784959793, + 0.020136920735239983, + 0.019131874665617943, + 0.020561477169394493, + 0.00771870743483305, + 0.016240103170275688, + 0.009966904297471046, + 0.009676615707576275, + 0.007077151909470558, + 0.0048660701140761375, + 0.003558120457455516, + 0.0034894561395049095, + 0.004493914544582367, + 0.004348110873252153, + 0.0030376091599464417, + 0.0039917100220918655, + 0.007675656583160162, + 0.006639642175287008, + 0.007050344254821539, + 0.007581240497529507, + 0.005007346626371145, + 0.006287583149969578, + 0.006488678976893425, + 0.006665397901087999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.04306214302778244 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a deserted train platform at night?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a young man with dark, slightly unkempt hair wearing a charcoal-gray hoodie and dark jeans?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the young man standing motionless with his hands tucked deep into his pockets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment illuminated by a single overhead lamp casting long, sharp shadows on weathered concrete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290732026100159 + ], + "opposite_logprob_tokens": [ + -2.2379074096679688 + ], + "logprob_reward": -0.11290732026100159, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290732026100159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18076026439666748, + 0.06400658190250397, + 0.04051758721470833, + 0.0647256150841713, + 0.11972030997276306, + 0.12201699614524841, + 0.19793523848056793, + 0.09344492852687836, + 0.07155501842498779, + 0.04531748592853546 + ], + "credit_map": [ + 0.027942050248384476, + 0.029284214600920677, + 0.020245486870408058, + 0.022529026493430138, + 0.01848388835787773, + 0.024864401668310165, + 0.021016091108322144, + 0.01639510504901409, + 0.010879190638661385, + 0.008044799789786339, + 0.008352779783308506, + 0.006663087289780378, + 0.0072810291312634945, + 0.010739796794950962, + 0.00670603895559907, + 0.005339855793863535, + 0.004890156909823418, + 0.004742589313536882, + 0.006817161571234465, + 0.004692950285971165, + 0.004238234832882881, + 0.0058382125571370125, + 0.0056390403769910336, + 0.0036592415999621153, + 0.010929174721240997, + 0.013063397258520126, + 0.007320637349039316, + 0.004169851075857878, + 0.008493329398334026, + 0.012001710012555122, + 0.005347210913896561, + 0.0034002987667918205, + 0.02430850639939308, + 0.025308387354016304, + 0.01741539128124714, + 0.010273885913193226, + 0.00916148629039526, + 0.016469275578856468, + 0.010098114609718323, + 0.0066852509044110775, + 0.025671757757663727, + 0.03244588151574135, + 0.015345302410423756, + 0.008743014186620712, + 0.008019325323402882, + 0.018648551777005196, + 0.006957603618502617, + 0.006185559090226889, + 0.034029096364974976, + 0.028660297393798828, + 0.027464454993605614, + 0.016750238835811615, + 0.028331752866506577, + 0.020635176450014114, + 0.028360959142446518, + 0.013703248463571072, + 0.014311490580439568, + 0.017854226753115654, + 0.012455309741199017, + 0.011893321759998798, + 0.007891927845776081, + 0.012125388719141483, + 0.009273971430957317, + 0.007639283314347267, + 0.00922902300953865, + 0.008100415579974651, + 0.0099975960329175, + 0.006066649220883846, + 0.0102913873270154, + 0.00822740700095892, + 0.009820714592933655, + 0.009821821004152298, + 0.005532945040613413, + 0.005818304605782032, + 0.005884528160095215, + 0.004835785366594791, + 0.0052572633139789104, + 0.005478025414049625, + 0.006819953676313162, + 0.005690678488463163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.11290732026100159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_6/index_tensor([7679, 7679])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_6/index_tensor([7679, 7679])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..910edd6010df03e5877af1dad96ae138737def03 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_6/index_tensor([7679, 7679])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "still, dramatic overhead close-up, captured in stunning 8K resolution and extreme slow motion, like a decisive moment from a World Cup match, focuses on the surface of fertile, dark, moist soil, slowly rising and cracking under a force rising from beneath the surface. Suddenly, with an unstoppable, primal force, a vibrant, emerald green sprout bursts through the earth, its two cotyledons curled tightly like clenched fists, ready to announce its arrival to the world. With its emergence, countless individual soil particles, from fine dust to tiny clods of earth, are violently hurled upward and outward, forming a spectacular, solidified cloud of explosions in the air. The rough texture and irregular shape of each particle are clearly visible in the ultra-high resolution. A strong beam of early morning sunlight slants into the frame, outlining the flying soil particles with golden hues and making the shimmering dewdrops covering the sprout sparkle like tiny diamonds, contrasting sharply against the deep earthy background. The entire scene is filled with a primitive sense of power and speed, transforming a microscopic process of life birth into a grand, tension-filled and shocking visual epic.", + "seed": 30334, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_6/index_7679_seed_30334_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.8196379453875124, + "vlm_logprob": -0.8196379453875124, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.8196379453875124, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00748058594763279 + ], + "opposite_logprob_tokens": [ + -5.882480621337891 + ], + "logprob_reward": -0.00748058594763279, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00748058594763279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1030406579375267, + 0.02439154125750065, + 0.042304329574108124, + 0.030457565560936928, + 0.5653260350227356, + 0.03705086559057236, + 0.0537848100066185, + 0.056011248379945755, + 0.06550370901823044, + 0.022129185497760773 + ], + "credit_map": [ + 0.013044487684965134, + 0.0078078885562717915, + 0.023229118436574936, + 0.00652953889220953, + 0.006783560384064913, + 0.028652982786297798, + 0.007907330058515072, + 0.009085754863917828, + 0.0024054988753050566, + 0.00285364524461329, + 0.004023024346679449, + 0.00374760664999485, + 0.002307126997038722, + 0.0033673506695777178, + 0.0034187398850917816, + 0.0022685506846755743, + 0.004881240427494049, + 0.005796667188405991, + 0.005908775608986616, + 0.0045768022537231445, + 0.0047294278629124165, + 0.007829728536307812, + 0.0048554399982094765, + 0.0037262493278831244, + 0.0037825556937605143, + 0.004122704267501831, + 0.004730834160000086, + 0.0028882459737360477, + 0.0034177228808403015, + 0.003198297694325447, + 0.003507049521431327, + 0.004810158628970385, + 0.1326281726360321, + 0.016220254823565483, + 0.11200816929340363, + 0.022632669657468796, + 0.016659127548336983, + 0.20466122031211853, + 0.03156440705060959, + 0.028952067717909813, + 0.0049272864125669, + 0.004968969617038965, + 0.004764512646943331, + 0.004145842976868153, + 0.005076679866760969, + 0.005424403119832277, + 0.0034874286502599716, + 0.004255746491253376, + 0.007273434195667505, + 0.006435575895011425, + 0.006940177641808987, + 0.00712420791387558, + 0.004307072144001722, + 0.009029306471347809, + 0.005528199020773172, + 0.007146842777729034, + 0.007062301971018314, + 0.00819320697337389, + 0.005423140712082386, + 0.008213017135858536, + 0.006996573880314827, + 0.005728818476200104, + 0.00856624823063612, + 0.005827946588397026, + 0.009151378646492958, + 0.007960527203977108, + 0.008256753906607628, + 0.0059934696182608604, + 0.009307359345257282, + 0.010236615315079689, + 0.007987253367900848, + 0.0066103567369282246, + 0.002335815690457821, + 0.0028581032529473305, + 0.0028817136771976948, + 0.003750741947442293, + 0.0025055240839719772, + 0.002478544833138585, + 0.0024154100101441145, + 0.002903335029259324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.00748058594763279 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005976306740194559 + ], + "opposite_logprob_tokens": [ + -5.13097620010376 + ], + "logprob_reward": -0.005976306740194559, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005976306740194559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10688735544681549, + 0.04860301315784454, + 0.03399677202105522, + 0.08902618288993835, + 0.40924131870269775, + 0.11739547550678253, + 0.04455755650997162, + 0.05056968703866005, + 0.0427706316113472, + 0.056951992213726044 + ], + "credit_map": [ + 0.01714513637125492, + 0.01966380886733532, + 0.015641195699572563, + 0.00791527796536684, + 0.011825024150311947, + 0.014452147297561169, + 0.010843425057828426, + 0.009401337243616581, + 0.005876653827726841, + 0.006828456185758114, + 0.012538298964500427, + 0.006349827162921429, + 0.0036543288733810186, + 0.005091813858598471, + 0.004427469335496426, + 0.003836164716631174, + 0.003179963445290923, + 0.007390845101326704, + 0.0046378388069570065, + 0.003751266747713089, + 0.0032369312830269337, + 0.004315637052059174, + 0.004139369819313288, + 0.0033449179027229548, + 0.01010425016283989, + 0.014662211760878563, + 0.02339022234082222, + 0.007323927711695433, + 0.004545255564153194, + 0.010016260668635368, + 0.00800285767763853, + 0.010981197468936443, + 0.05252194404602051, + 0.012644887901842594, + 0.1650920957326889, + 0.026805348694324493, + 0.011927437968552113, + 0.087370865046978, + 0.02788734994828701, + 0.02499140240252018, + 0.01589205674827099, + 0.018607601523399353, + 0.01812819205224514, + 0.011861857026815414, + 0.00691835954785347, + 0.02005917578935623, + 0.008712860755622387, + 0.017215358093380928, + 0.006468311417847872, + 0.005941420327872038, + 0.004393556155264378, + 0.003941788338124752, + 0.005432932171970606, + 0.010212183929979801, + 0.004243853967636824, + 0.003923512063920498, + 0.0067428299225866795, + 0.005134037230163813, + 0.007239378988742828, + 0.007589451968669891, + 0.0032848783303052187, + 0.009189799427986145, + 0.004551965743303299, + 0.006837345659732819, + 0.004484747536480427, + 0.0038603516295552254, + 0.008484946563839912, + 0.005640250165015459, + 0.005126248113811016, + 0.006428021937608719, + 0.0039850021712481976, + 0.004761063493788242, + 0.01106419786810875, + 0.010157582350075245, + 0.008274879306554794, + 0.005419817753136158, + 0.003770010080188513, + 0.007374139968305826, + 0.003501854371279478, + 0.007389516104012728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.005976306740194559 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983338013291359 + ], + "opposite_logprob_tokens": [ + -3.5298333168029785 + ], + "logprob_reward": -0.02983338013291359, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983338013291359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052797820419073105, + 0.04448188841342926, + 0.04703832417726517, + 0.025112738832831383, + 0.6603306531906128, + 0.028107061982154846, + 0.04620172083377838, + 0.028907600790262222, + 0.05155186355113983, + 0.015470344573259354 + ], + "credit_map": [ + 0.0070845806039869785, + 0.0034168888814747334, + 0.011267848312854767, + 0.0036903563886880875, + 0.003657132387161255, + 0.011698308400809765, + 0.005574780981987715, + 0.006407923065125942, + 0.004378268960863352, + 0.005444810725748539, + 0.006896478123962879, + 0.005475315731018782, + 0.0033780084922909737, + 0.007886872626841068, + 0.006722395308315754, + 0.004299739841371775, + 0.00423486577346921, + 0.00963612087070942, + 0.005957991350442171, + 0.004591675940901041, + 0.005336009431630373, + 0.007609822321683168, + 0.005655491258949041, + 0.004016343504190445, + 0.0037024025805294514, + 0.003850962733849883, + 0.004645973909646273, + 0.0023698986042290926, + 0.002189550781622529, + 0.002461867406964302, + 0.00273700850084424, + 0.003155072685331106, + 0.147740438580513, + 0.015621891245245934, + 0.14753831923007965, + 0.027792325243353844, + 0.0166412815451622, + 0.236146941781044, + 0.0347263403236866, + 0.03412310779094696, + 0.0030613024719059467, + 0.0038615281227976084, + 0.004429066553711891, + 0.004369593225419521, + 0.002054263139143586, + 0.0038604664150625467, + 0.002355883363634348, + 0.004114960320293903, + 0.006507887039333582, + 0.003523788182064891, + 0.005375168286263943, + 0.003875934984534979, + 0.004105893429368734, + 0.012596657499670982, + 0.005208048969507217, + 0.00500834546983242, + 0.0049742478877305984, + 0.002996360883116722, + 0.003336218185722828, + 0.0029457516502588987, + 0.0033134364057332277, + 0.004814435262233019, + 0.0027652571443468332, + 0.0037618933711200953, + 0.0060032084584236145, + 0.0069975401274859905, + 0.005996534135192633, + 0.006541936658322811, + 0.008329955860972404, + 0.007427961099892855, + 0.005904887802898884, + 0.00434983940795064, + 0.0016880186740309, + 0.001983905676752329, + 0.0027240952476859093, + 0.0018848233157768846, + 0.0018028273480013013, + 0.0016073823207989335, + 0.0015514022670686245, + 0.0022278898395597935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.02983338013291359 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -3.7732739448547363 + ], + "opposite_logprob_tokens": [ + -0.023273982107639313 + ], + "logprob_reward": -3.7732739448547363, + "yes_no_margin_reward": -3.75, + "token_credit_reward": -3.7732739448547363, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0402991846203804, + 0.017714831978082657, + 0.014170505106449127, + 0.018740003928542137, + 0.7984788417816162, + 0.012354154139757156, + 0.01649363897740841, + 0.017995523288846016, + 0.011874140240252018, + 0.05187918245792389 + ], + "credit_map": [ + 0.005883379373699427, + 0.006801959127187729, + 0.006114973220974207, + 0.0028313950169831514, + 0.0040307664312422276, + 0.005537050310522318, + 0.005190195050090551, + 0.00390946539118886, + 0.0017343676881864667, + 0.0027382983826100826, + 0.0040330830961465836, + 0.0022082957439124584, + 0.001536440453492105, + 0.0016372593818232417, + 0.002380902413278818, + 0.001446184003725648, + 0.0016941732028499246, + 0.0015360622201114893, + 0.0016724702436476946, + 0.001266133040189743, + 0.00234474940225482, + 0.0027270044665783644, + 0.0016921957721933722, + 0.0012377158273011446, + 0.0023590659257024527, + 0.0019917574245482683, + 0.004299105145037174, + 0.001635287539102137, + 0.00161739659961313, + 0.002392203314229846, + 0.002039042767137289, + 0.002406145678833127, + 0.32022786140441895, + 0.025465315207839012, + 0.0636720359325409, + 0.027138637378811836, + 0.014100461266934872, + 0.25663647055625916, + 0.04854419827461243, + 0.0426938496530056, + 0.0015610986156389117, + 0.0014703378546983004, + 0.0015910594956949353, + 0.0013280796119943261, + 0.0012354624923318624, + 0.0023119011893868446, + 0.0013679477851837873, + 0.0014882672112435102, + 0.0023157449904829264, + 0.0013447330566123128, + 0.002275872742757201, + 0.002339613623917103, + 0.0010715839453041553, + 0.003456811886280775, + 0.0016557129565626383, + 0.002033565193414688, + 0.002873958321288228, + 0.001811739057302475, + 0.0024315740447491407, + 0.002256924519315362, + 0.0013654509093612432, + 0.0033165281638503075, + 0.001594213885255158, + 0.0023451342713087797, + 0.0012975012650713325, + 0.0015182069037109613, + 0.0019614268094301224, + 0.0012778794625774026, + 0.001525979139842093, + 0.0017071288311854005, + 0.0011341273784637451, + 0.0014518896350637078, + 0.008445548824965954, + 0.010084683075547218, + 0.005659560672938824, + 0.0069076018407940865, + 0.00428732018917799, + 0.006138000171631575, + 0.0045263865031301975, + 0.005830086302012205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -3.7732739448547363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28162550926208496 + ], + "opposite_logprob_tokens": [ + -1.406625509262085 + ], + "logprob_reward": -0.28162550926208496, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28162550926208496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04434671252965927, + 0.04507729411125183, + 0.021381696686148643, + 0.05293453857302666, + 0.6476330161094666, + 0.016939429566264153, + 0.024440061300992966, + 0.026269854977726936, + 0.020969459787011147, + 0.10000790655612946 + ], + "credit_map": [ + 0.007190093398094177, + 0.005736891645938158, + 0.007143673952668905, + 0.003421686589717865, + 0.0037654894404113293, + 0.006080912426114082, + 0.005798668134957552, + 0.00520930252969265, + 0.0050134919583797455, + 0.006363441701978445, + 0.006959811318665743, + 0.005445396527647972, + 0.0034730739425867796, + 0.01064709760248661, + 0.003855894086882472, + 0.003319092560559511, + 0.0018806892912834883, + 0.0041369604878127575, + 0.0029017170891165733, + 0.002178453141823411, + 0.002508978359401226, + 0.0035667584743350744, + 0.0021355929784476757, + 0.002072548493742943, + 0.006798258516937494, + 0.0068106758408248425, + 0.013100355863571167, + 0.0039278180338442326, + 0.0039560082368552685, + 0.007131617981940508, + 0.005029850173741579, + 0.00617996184155345, + 0.25221580266952515, + 0.021434232592582703, + 0.10832412540912628, + 0.025378398597240448, + 0.012325995601713657, + 0.15245507657527924, + 0.04119769483804703, + 0.0343017540872097, + 0.0019811447709798813, + 0.002334120450541377, + 0.0024410446640104055, + 0.001475884229876101, + 0.0019950729329138994, + 0.002908896654844284, + 0.0018665691604837775, + 0.0019366990309208632, + 0.0030062696896493435, + 0.002539399778470397, + 0.0033048673067241907, + 0.002852324629202485, + 0.00268420297652483, + 0.004253507126122713, + 0.0031097119208425283, + 0.002689782064408064, + 0.00418044812977314, + 0.004102903418242931, + 0.0027134581468999386, + 0.002755238674581051, + 0.0025586686097085476, + 0.003298576921224594, + 0.003409924451261759, + 0.0032506391871720552, + 0.002437855815514922, + 0.0021847097668796778, + 0.0036742386873811483, + 0.003308893647044897, + 0.002832893980666995, + 0.0023648804053664207, + 0.002261398360133171, + 0.0019045926164835691, + 0.01526328269392252, + 0.016000673174858093, + 0.01436040922999382, + 0.014195796102285385, + 0.0067637404426932335, + 0.011572910472750664, + 0.010380267165601254, + 0.011470841243863106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.28162550926208496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "still, dramatic overhead close-up, captured in stunning 8K resolution and extreme slow motion, like a decisive moment from a World Cup match, focuses on the surface of fertile, dark, moist soil, slowly rising and cracking under a force rising from beneath the surface. Suddenly, with an unstoppable, primal force, a vibrant, emerald green sprout bursts through the earth, its two cotyledons curled tightly like clenched fists, ready to announce its arrival to the world. With its emergence, countless individual soil particles, from fine dust to tiny clods of earth, are violently hurled upward and outward, forming a spectacular, solidified cloud of explosions in the air. The rough texture and irregular shape of each particle are clearly visible in the ultra-high resolution. A strong beam of early morning sunlight slants into the frame, outlining the flying soil particles with golden hues and making the shimmering dewdrops covering the sprout sparkle like tiny diamonds, contrasting sharply against the deep earthy background. The entire scene is filled with a primitive sense of power and speed, transforming a microscopic process of life birth into a grand, tension-filled and shocking visual epic.", + "seed": 30335, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_6/index_7679_seed_30335_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.5100459929555654, + "vlm_logprob": -0.5100459929555654, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.5100459929555654, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013598780147731304 + ], + "opposite_logprob_tokens": [ + -4.888598918914795 + ], + "logprob_reward": -0.013598780147731304, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.013598780147731304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0941353291273117, + 0.05440075322985649, + 0.03770826384425163, + 0.0390576533973217, + 0.0748210921883583, + 0.03500067815184593, + 0.05569063499569893, + 0.17856009304523468, + 0.3509794771671295, + 0.07964608073234558 + ], + "credit_map": [ + 0.014431419782340527, + 0.01685788854956627, + 0.024508561939001083, + 0.011694411747157574, + 0.006207631900906563, + 0.00798074807971716, + 0.006439224351197481, + 0.006015428341925144, + 0.006619954481720924, + 0.005505294073373079, + 0.007537669502198696, + 0.009958552196621895, + 0.0057266647927463055, + 0.006690893787890673, + 0.005804377142339945, + 0.006557343993335962, + 0.005821234080940485, + 0.0033075627870857716, + 0.007085652090609074, + 0.0032328455708920956, + 0.0022633629851043224, + 0.005463668145239353, + 0.003788936650380492, + 0.006744997575879097, + 0.004684611223638058, + 0.006391986273229122, + 0.00586197804659605, + 0.005100221838802099, + 0.006029329262673855, + 0.0041116210632026196, + 0.004695390351116657, + 0.002182512078434229, + 0.008240467868745327, + 0.004784089513123035, + 0.009123659692704678, + 0.007411566097289324, + 0.005557163152843714, + 0.01741720736026764, + 0.01243298128247261, + 0.009853948839008808, + 0.004436789080500603, + 0.0026305965147912502, + 0.005340747069567442, + 0.004097124561667442, + 0.0038512777537107468, + 0.004575433675199747, + 0.006214840803295374, + 0.0038538645021617413, + 0.007152243983000517, + 0.005303874146193266, + 0.007055320776998997, + 0.007432012353092432, + 0.007311226800084114, + 0.009259979240596294, + 0.006852803286164999, + 0.005323170218616724, + 0.021433044224977493, + 0.015565126203000546, + 0.022391116246581078, + 0.021771080791950226, + 0.01773405447602272, + 0.03910831734538078, + 0.019091716036200523, + 0.021465647965669632, + 0.049308694899082184, + 0.033576786518096924, + 0.05030176043510437, + 0.07377006113529205, + 0.03061489574611187, + 0.026118559762835503, + 0.06764144450426102, + 0.019647255539894104, + 0.01586543209850788, + 0.00577534968033433, + 0.012857098132371902, + 0.007876968942582607, + 0.005896314047276974, + 0.006126813590526581, + 0.00977405346930027, + 0.015474037267267704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.013598780147731304 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009805607609450817 + ], + "opposite_logprob_tokens": [ + -4.634805679321289 + ], + "logprob_reward": -0.009805607609450817, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009805607609450817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2556939423084259, + 0.16393567621707916, + 0.04784572124481201, + 0.03456088900566101, + 0.11202484369277954, + 0.06062868982553482, + 0.05542970821261406, + 0.033942993730306625, + 0.09037844836711884, + 0.145559161901474 + ], + "credit_map": [ + 0.020753711462020874, + 0.07415108382701874, + 0.06641429662704468, + 0.023268379271030426, + 0.016113076359033585, + 0.021044880151748657, + 0.01870979182422161, + 0.015238688327372074, + 0.02314227633178234, + 0.014257743023335934, + 0.02036799117922783, + 0.025485826656222343, + 0.013111271895468235, + 0.019985070452094078, + 0.014606062322854996, + 0.032979417592287064, + 0.008075849153101444, + 0.004487453959882259, + 0.007450130768120289, + 0.0038138043601065874, + 0.00470008933916688, + 0.00831528939306736, + 0.004425543360412121, + 0.0065775541588664055, + 0.0037929026875644922, + 0.0065824054181575775, + 0.0034023893531411886, + 0.003267646534368396, + 0.004962171893566847, + 0.003995340783149004, + 0.005151254590600729, + 0.0034067716915160418, + 0.01692589558660984, + 0.008099878206849098, + 0.02090495638549328, + 0.012338044121861458, + 0.008449218235909939, + 0.01708230748772621, + 0.014038298279047012, + 0.014186236076056957, + 0.005320650525391102, + 0.006713625509291887, + 0.010294483043253422, + 0.006047171074897051, + 0.008485011756420135, + 0.011473619379103184, + 0.008002624846994877, + 0.004291498567909002, + 0.00734354555606842, + 0.0057640280574560165, + 0.00980682298541069, + 0.007180023938417435, + 0.006892653647810221, + 0.007509193383157253, + 0.0055619231425225735, + 0.005371508188545704, + 0.004438396077603102, + 0.0028919032774865627, + 0.005852971225976944, + 0.004332053940743208, + 0.0024007228203117847, + 0.005357066169381142, + 0.003996043000370264, + 0.00467383349314332, + 0.012976134195923805, + 0.00668939808383584, + 0.022759953513741493, + 0.013934578746557236, + 0.00618678517639637, + 0.007574667688459158, + 0.010821723379194736, + 0.009435196407139301, + 0.0221351720392704, + 0.008759957738220692, + 0.023291483521461487, + 0.018971402198076248, + 0.009525347501039505, + 0.019533364102244377, + 0.01789240725338459, + 0.025450019165873528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.009805607609450817 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016085729002952576 + ], + "opposite_logprob_tokens": [ + -4.141085624694824 + ], + "logprob_reward": -0.016085729002952576, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016085729002952576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10581428557634354, + 0.11689385771751404, + 0.07634688168764114, + 0.04911493882536888, + 0.06425367295742035, + 0.07194693386554718, + 0.10042054951190948, + 0.20045708119869232, + 0.1605268120765686, + 0.05422497168183327 + ], + "credit_map": [ + 0.00859557744115591, + 0.03126855194568634, + 0.01298945676535368, + 0.011426996439695358, + 0.01197199709713459, + 0.010043183341622353, + 0.011551037430763245, + 0.007967486046254635, + 0.012516850605607033, + 0.013953668996691704, + 0.01817859709262848, + 0.014266381971538067, + 0.01169434655457735, + 0.02365017682313919, + 0.010865549556910992, + 0.011768283322453499, + 0.016292132437229156, + 0.00681262044236064, + 0.010071840137243271, + 0.006694714538753033, + 0.0066864266991615295, + 0.007925505749881268, + 0.008296818472445011, + 0.01356683112680912, + 0.005484726279973984, + 0.008817746303975582, + 0.006050923839211464, + 0.005008188541978598, + 0.006920858286321163, + 0.005541871301829815, + 0.006912455894052982, + 0.0043781716376543045, + 0.005820910912007093, + 0.005214906297624111, + 0.008752865716814995, + 0.006839127745479345, + 0.005564897786825895, + 0.011701474897563457, + 0.01276269368827343, + 0.007596796378493309, + 0.008239107206463814, + 0.0059493230655789375, + 0.010600586421787739, + 0.007544808555394411, + 0.007422857917845249, + 0.012378256767988205, + 0.012253742665052414, + 0.007558249402791262, + 0.011894476599991322, + 0.0076211863197386265, + 0.013194539584219456, + 0.013420050032436848, + 0.010293090716004372, + 0.019094299525022507, + 0.012396321631968021, + 0.01250658743083477, + 0.02266213484108448, + 0.02146337367594242, + 0.023259609937667847, + 0.02122940868139267, + 0.023157810792326927, + 0.04280642792582512, + 0.021325385197997093, + 0.024552930146455765, + 0.028157182037830353, + 0.01546586025506258, + 0.02073102444410324, + 0.034710437059402466, + 0.012544309720396996, + 0.012682314962148666, + 0.02404063567519188, + 0.012195047922432423, + 0.006412841379642487, + 0.004389662761241198, + 0.010356075130403042, + 0.009566365741193295, + 0.003995269536972046, + 0.004997763782739639, + 0.005104509647935629, + 0.009402484633028507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.016085729002952576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.350275993347168 + ], + "opposite_logprob_tokens": [ + -0.10027603060007095 + ], + "logprob_reward": -2.350275993347168, + "yes_no_margin_reward": -2.25, + "token_credit_reward": -2.350275993347168, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11867181956768036, + 0.3533738851547241, + 0.0193789079785347, + 0.09938537329435349, + 0.029246963560581207, + 0.08883170038461685, + 0.023146433755755424, + 0.053994279354810715, + 0.16126875579357147, + 0.05270185321569443 + ], + "credit_map": [ + 0.011299682781100273, + 0.04096481204032898, + 0.021204471588134766, + 0.007804608438163996, + 0.008299804292619228, + 0.015924489125609398, + 0.006019750144332647, + 0.007154208607971668, + 0.04229087010025978, + 0.041219718754291534, + 0.052421581000089645, + 0.05825149267911911, + 0.036008916795253754, + 0.06041499227285385, + 0.030418775975704193, + 0.03234757110476494, + 0.002291299169883132, + 0.0018816010560840368, + 0.002984684193506837, + 0.0019801657181233168, + 0.0014793598093092442, + 0.002633708994835615, + 0.003478533122688532, + 0.0026495580095797777, + 0.012782715260982513, + 0.01658252812922001, + 0.012120998464524746, + 0.013480874709784985, + 0.011836064979434013, + 0.010258936323225498, + 0.01197524182498455, + 0.010348021984100342, + 0.004203423857688904, + 0.0017964128637686372, + 0.0037369432393461466, + 0.0028233258053660393, + 0.002153565874323249, + 0.006919453386217356, + 0.0028459164313971996, + 0.004767924547195435, + 0.008182871155440807, + 0.00924699753522873, + 0.013761429116129875, + 0.008948798291385174, + 0.011836559511721134, + 0.015139566734433174, + 0.015059350058436394, + 0.006656138226389885, + 0.003190886927768588, + 0.001195517717860639, + 0.0029171560890972614, + 0.002434497233480215, + 0.0016664506401866674, + 0.005640198476612568, + 0.0028825104236602783, + 0.003219219157472253, + 0.006451699882745743, + 0.005120859481394291, + 0.006515929009765387, + 0.006808220408856869, + 0.005722634959965944, + 0.01106264628469944, + 0.006151903420686722, + 0.006160387769341469, + 0.02158227376639843, + 0.014820707961916924, + 0.02199145220220089, + 0.033878348767757416, + 0.014627569355070591, + 0.011778593994677067, + 0.03310048207640648, + 0.009489347226917744, + 0.008174595423042774, + 0.004496970679610968, + 0.00947806891053915, + 0.0056283301673829556, + 0.0046709878370165825, + 0.005200459621846676, + 0.006198929622769356, + 0.008853517472743988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -2.350275993347168 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604638546705246 + ], + "opposite_logprob_tokens": [ + -1.910463809967041 + ], + "logprob_reward": -0.1604638546705246, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604638546705246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11904049664735794, + 0.3467554748058319, + 0.033043500036001205, + 0.04947368800640106, + 0.04826589301228523, + 0.09968343377113342, + 0.03283379226922989, + 0.09585965424776077, + 0.10194839537143707, + 0.07309568673372269 + ], + "credit_map": [ + 0.017032917588949203, + 0.015951581299304962, + 0.02910505048930645, + 0.01389991957694292, + 0.011123660951852798, + 0.014760706573724747, + 0.010008497163653374, + 0.007158162072300911, + 0.04514629766345024, + 0.03847741335630417, + 0.04878222569823265, + 0.056281015276908875, + 0.036287758499383926, + 0.057731159031391144, + 0.03014683723449707, + 0.03390275314450264, + 0.003972450736910105, + 0.0033001317642629147, + 0.004521884489804506, + 0.0037927222438156605, + 0.0038357048761099577, + 0.004243749193847179, + 0.004451103508472443, + 0.00492575578391552, + 0.007108476012945175, + 0.008034578524529934, + 0.006709511391818523, + 0.0045226686634123325, + 0.00620329799130559, + 0.006298841908574104, + 0.0068660578690469265, + 0.0037302589043974876, + 0.005137114319950342, + 0.005368213634938002, + 0.006195446942001581, + 0.005575371906161308, + 0.006623155903071165, + 0.00784168392419815, + 0.006186064798384905, + 0.0053388383239507675, + 0.010285284370183945, + 0.010308272205293179, + 0.013785358518362045, + 0.010421761311590672, + 0.012912167236208916, + 0.015117787756025791, + 0.017746834084391594, + 0.009105968289077282, + 0.0049837855622172356, + 0.0026351327542215586, + 0.0033916968386620283, + 0.0041850837878882885, + 0.0031585467513650656, + 0.00676171388477087, + 0.0038962438702583313, + 0.0038215916138142347, + 0.011543883010745049, + 0.008848680183291435, + 0.01135121751576662, + 0.01255082618445158, + 0.01000922080129385, + 0.019605370238423347, + 0.011219615116715431, + 0.010730842128396034, + 0.012903247959911823, + 0.00776210380718112, + 0.016181062906980515, + 0.024939153343439102, + 0.0075993649661540985, + 0.009709564968943596, + 0.015358490869402885, + 0.007495402358472347, + 0.012951286509633064, + 0.005460500251501799, + 0.010784143581986427, + 0.015509740449488163, + 0.00463856291025877, + 0.00572326872497797, + 0.009745118208229542, + 0.008283066563308239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.1604638546705246 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "still, dramatic overhead close-up, captured in stunning 8K resolution and extreme slow motion, like a decisive moment from a World Cup match, focuses on the surface of fertile, dark, moist soil, slowly rising and cracking under a force rising from beneath the surface. Suddenly, with an unstoppable, primal force, a vibrant, emerald green sprout bursts through the earth, its two cotyledons curled tightly like clenched fists, ready to announce its arrival to the world. With its emergence, countless individual soil particles, from fine dust to tiny clods of earth, are violently hurled upward and outward, forming a spectacular, solidified cloud of explosions in the air. The rough texture and irregular shape of each particle are clearly visible in the ultra-high resolution. A strong beam of early morning sunlight slants into the frame, outlining the flying soil particles with golden hues and making the shimmering dewdrops covering the sprout sparkle like tiny diamonds, contrasting sharply against the deep earthy background. The entire scene is filled with a primitive sense of power and speed, transforming a microscopic process of life birth into a grand, tension-filled and shocking visual epic.", + "seed": 30336, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_6/index_7679_seed_30336_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.46192993139848115, + "vlm_logprob": -0.46192993139848115, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.46192993139848115, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006597877945750952 + ], + "opposite_logprob_tokens": [ + -6.006597995758057 + ], + "logprob_reward": -0.006597877945750952, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.006597877945750952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2186935693025589, + 0.10937094688415527, + 0.03192323446273804, + 0.1251448392868042, + 0.052577994763851166, + 0.040420521050691605, + 0.15486879646778107, + 0.1139773577451706, + 0.135240375995636, + 0.01778233051300049 + ], + "credit_map": [ + 0.050054024904966354, + 0.008196634240448475, + 0.01053430512547493, + 0.00982504803687334, + 0.08335041254758835, + 0.009471245110034943, + 0.015166684985160828, + 0.032095227390527725, + 0.008026930503547192, + 0.005521811544895172, + 0.007711400743573904, + 0.00616883672773838, + 0.03749334439635277, + 0.004679752979427576, + 0.009256944991648197, + 0.030511926859617233, + 0.0042777410708367825, + 0.0038992748595774174, + 0.003982706926763058, + 0.00549244275316596, + 0.004843465052545071, + 0.0024622248020023108, + 0.0037072342820465565, + 0.0032581444829702377, + 0.022496335208415985, + 0.00530331302434206, + 0.010211816057562828, + 0.004738072864711285, + 0.008061833679676056, + 0.006883004680275917, + 0.025944339111447334, + 0.04150611534714699, + 0.006997790187597275, + 0.004103776533156633, + 0.006467440631240606, + 0.006573801394551992, + 0.00488660391420126, + 0.010254447348415852, + 0.007533453404903412, + 0.005760682746767998, + 0.0047868164256215096, + 0.004077829420566559, + 0.005158733576536179, + 0.007091629318892956, + 0.0032138426322489977, + 0.004274878650903702, + 0.007755001075565815, + 0.004061789717525244, + 0.017219215631484985, + 0.015332126058638096, + 0.02290661819279194, + 0.02507607638835907, + 0.01605839468538761, + 0.016397712752223015, + 0.027292588725686073, + 0.014586061239242554, + 0.012756549753248692, + 0.011599981226027012, + 0.015572812408208847, + 0.013620360754430294, + 0.01382317766547203, + 0.017532596364617348, + 0.015033756382763386, + 0.014038131572306156, + 0.01203785091638565, + 0.010274623520672321, + 0.017612142488360405, + 0.019474029541015625, + 0.01676187291741371, + 0.02225371263921261, + 0.022289246320724487, + 0.014536905102431774, + 0.0014691530959680676, + 0.0014390151482075453, + 0.0021833712235093117, + 0.002711124252527952, + 0.0018677701009437442, + 0.003972080536186695, + 0.002402689540758729, + 0.0017371257999911904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.006597877945750952 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009813634678721428 + ], + "opposite_logprob_tokens": [ + -4.6348137855529785 + ], + "logprob_reward": -0.009813634678721428, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009813634678721428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22365392744541168, + 0.0628582015633583, + 0.03281811624765396, + 0.21528774499893188, + 0.13899986445903778, + 0.04795972257852554, + 0.0690184235572815, + 0.06431901454925537, + 0.0934411883354187, + 0.05164378881454468 + ], + "credit_map": [ + 0.03862820938229561, + 0.016116591170430183, + 0.0216684527695179, + 0.02090233936905861, + 0.059646546840667725, + 0.014481834135949612, + 0.022566264495253563, + 0.029643690213561058, + 0.010896177031099796, + 0.006178930401802063, + 0.010993271134793758, + 0.007149762474000454, + 0.006674367468804121, + 0.005920396652072668, + 0.007944351062178612, + 0.007100946269929409, + 0.005687974393367767, + 0.003103225491940975, + 0.0025262769777327776, + 0.0033513254020363092, + 0.0029591734055429697, + 0.0031149107962846756, + 0.008533147163689137, + 0.0035420823842287064, + 0.0689532458782196, + 0.014763605780899525, + 0.021389257162809372, + 0.009577576071023941, + 0.010045075789093971, + 0.017455462366342545, + 0.030965574085712433, + 0.04213795065879822, + 0.01901840977370739, + 0.013106781989336014, + 0.01585737243294716, + 0.018634486943483353, + 0.014011736959218979, + 0.021245311945676804, + 0.020528066903352737, + 0.016597697511315346, + 0.006188098806887865, + 0.004885062575340271, + 0.006622994784265757, + 0.006127985659986734, + 0.004288742318749428, + 0.006977098062634468, + 0.0057281022891402245, + 0.007141638547182083, + 0.006594937760382891, + 0.007028651423752308, + 0.007604830898344517, + 0.01507192850112915, + 0.007137469481676817, + 0.00621934374794364, + 0.012459474615752697, + 0.0069017861969769, + 0.005860303994268179, + 0.006453987676650286, + 0.006918896455317736, + 0.007238489575684071, + 0.0058236801996827126, + 0.009348698891699314, + 0.008876727893948555, + 0.013798235915601254, + 0.0068212090991437435, + 0.0076400465331971645, + 0.017268063500523567, + 0.011959428898990154, + 0.011591360904276371, + 0.008747290819883347, + 0.016184775158762932, + 0.013229007832705975, + 0.0060120453126728535, + 0.004515795968472958, + 0.008826402015984058, + 0.006284493021667004, + 0.0036860276013612747, + 0.009166749194264412, + 0.006373818963766098, + 0.006778459995985031 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.009813634678721428 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331998571753502 + ], + "opposite_logprob_tokens": [ + -3.773319959640503 + ], + "logprob_reward": -0.02331998571753502, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331998571753502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14879682660102844, + 0.15784268081188202, + 0.03882487118244171, + 0.32063862681388855, + 0.08937419950962067, + 0.03333878889679909, + 0.05053705722093582, + 0.041467953473329544, + 0.0880952700972557, + 0.031083717942237854 + ], + "credit_map": [ + 0.01837891712784767, + 0.006398072931915522, + 0.008847124874591827, + 0.012615677900612354, + 0.05009474605321884, + 0.006560313981026411, + 0.01780545525252819, + 0.028096526861190796, + 0.019918348640203476, + 0.011974966153502464, + 0.020372210070490837, + 0.025960031896829605, + 0.02417682111263275, + 0.010736741125583649, + 0.02288287691771984, + 0.021820681169629097, + 0.005966059863567352, + 0.003810031106695533, + 0.005167088005691767, + 0.004554476588964462, + 0.005219507962465286, + 0.003522467566654086, + 0.007261672988533974, + 0.003323568496853113, + 0.042828913778066635, + 0.020007967948913574, + 0.025966838002204895, + 0.01589713618159294, + 0.020115196704864502, + 0.02202887088060379, + 0.07530748099088669, + 0.09848622232675552, + 0.007668771781027317, + 0.012918547727167606, + 0.007050693966448307, + 0.01613914594054222, + 0.01425094809383154, + 0.0048917969688773155, + 0.017404960468411446, + 0.009049337357282639, + 0.0030734101310372353, + 0.004623347893357277, + 0.004943635780364275, + 0.004990328103303909, + 0.004547866526991129, + 0.0028104884549975395, + 0.005607889033854008, + 0.0027418239042162895, + 0.003410125384107232, + 0.003629254875704646, + 0.0034065935760736465, + 0.01705782115459442, + 0.0035662036389112473, + 0.0042002541013062, + 0.010812163352966309, + 0.0044546425342559814, + 0.004463006742298603, + 0.005660845898091793, + 0.0046975817531347275, + 0.004544280469417572, + 0.006208390463143587, + 0.005760343745350838, + 0.005487707443535328, + 0.00464579788967967, + 0.007704524789005518, + 0.008170224726200104, + 0.010851409286260605, + 0.011305284686386585, + 0.012869262136518955, + 0.010183983482420444, + 0.016711384057998657, + 0.01029920019209385, + 0.0023921197280287743, + 0.002695837989449501, + 0.004082479048520327, + 0.003618147922679782, + 0.004903439898043871, + 0.0028859595768153667, + 0.006125887390226126, + 0.004379846155643463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.02331998571753502 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.127023458480835 + ], + "opposite_logprob_tokens": [ + -0.12702348828315735 + ], + "logprob_reward": -2.127023458480835, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.127023458480835, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3341655135154724, + 0.13966546952724457, + 0.06251592189073563, + 0.10421469807624817, + 0.07998839020729065, + 0.06531573832035065, + 0.07372059673070908, + 0.05687067657709122, + 0.06052142009139061, + 0.023021547123789787 + ], + "credit_map": [ + 0.046888720244169235, + 0.01733381301164627, + 0.013318910263478756, + 0.016956808045506477, + 0.13493488729000092, + 0.015834301710128784, + 0.019187692552804947, + 0.0697103962302208, + 0.022541096433997154, + 0.011860405094921589, + 0.022773556411266327, + 0.020918071269989014, + 0.014694394543766975, + 0.012648052535951138, + 0.018832137808203697, + 0.0153977544978261, + 0.011333146132528782, + 0.005358901806175709, + 0.006765465717762709, + 0.0054817283526062965, + 0.006819692440330982, + 0.0057252803817391396, + 0.015228122472763062, + 0.005803586449474096, + 0.013219363987445831, + 0.009515959769487381, + 0.008261285722255707, + 0.008877797052264214, + 0.01328123640269041, + 0.012662697583436966, + 0.016490910202264786, + 0.0219054464250803, + 0.011369548738002777, + 0.0072009991854429245, + 0.006894290447235107, + 0.009815093129873276, + 0.007399229798465967, + 0.010131645947694778, + 0.017351962625980377, + 0.009825619868934155, + 0.008527361787855625, + 0.007079560775309801, + 0.00474251713603735, + 0.009301102720201015, + 0.008766709826886654, + 0.010579409077763557, + 0.009767280891537666, + 0.006551791448146105, + 0.006052748300135136, + 0.007291541434824467, + 0.007442429196089506, + 0.018031088635325432, + 0.007535084616392851, + 0.006598741747438908, + 0.01506093330681324, + 0.005708036944270134, + 0.005760339554399252, + 0.005071532446891069, + 0.007205968257039785, + 0.0068474141880869865, + 0.005724246613681316, + 0.010068576782941818, + 0.007902026176452637, + 0.00829057116061449, + 0.006011964753270149, + 0.005535244010388851, + 0.006955403834581375, + 0.008151429705321789, + 0.00846534688025713, + 0.009852927178144455, + 0.009312600828707218, + 0.006236506626009941, + 0.002948248293250799, + 0.001989170676097274, + 0.0021471246145665646, + 0.0035368052776902914, + 0.002991192974150181, + 0.0034619539510458708, + 0.0037919313181191683, + 0.0021551207173615694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -2.127023458480835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428947001695633 + ], + "opposite_logprob_tokens": [ + -2.017894744873047 + ], + "logprob_reward": -0.1428947001695633, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428947001695633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5438380241394043, + 0.040231745690107346, + 0.03648785874247551, + 0.050858937203884125, + 0.022526411339640617, + 0.04053061082959175, + 0.1578080803155899, + 0.02784671075642109, + 0.06296644359827042, + 0.0169052816927433 + ], + "credit_map": [ + 0.07626722753047943, + 0.025748534128069878, + 0.01881198212504387, + 0.025341005995869637, + 0.22412006556987762, + 0.022940784692764282, + 0.03291289135813713, + 0.11769548803567886, + 0.005267161875963211, + 0.004348285496234894, + 0.006050639785826206, + 0.0046958536840975285, + 0.0052164769731462, + 0.003715074621140957, + 0.004720865748822689, + 0.006217387970536947, + 0.005182662978768349, + 0.003517954610288143, + 0.003804190782830119, + 0.003767258021980524, + 0.00424879789352417, + 0.0027136816643178463, + 0.009265990927815437, + 0.003987319301813841, + 0.01348593458533287, + 0.0037575571332126856, + 0.004958579316735268, + 0.002523246919736266, + 0.0035286827478557825, + 0.006629305426031351, + 0.006674312055110931, + 0.009301316924393177, + 0.002275830367580056, + 0.002605895511806011, + 0.00329286209307611, + 0.003315799869596958, + 0.0033299995120614767, + 0.0018994942074641585, + 0.0031325591262429953, + 0.0026739686727523804, + 0.003430294571444392, + 0.005459477659314871, + 0.004748268518596888, + 0.00843819696456194, + 0.004482042510062456, + 0.003364905482158065, + 0.005921120289713144, + 0.004686298780143261, + 0.01747179590165615, + 0.009822840802371502, + 0.014964884147047997, + 0.026845157146453857, + 0.014166026376187801, + 0.03546299412846565, + 0.022759277373552322, + 0.016315095126628876, + 0.0029585114680230618, + 0.0025287936441600323, + 0.0024241262581199408, + 0.0036039315164089203, + 0.0035218484699726105, + 0.004216646309942007, + 0.00462577398866415, + 0.003967075143009424, + 0.004934917204082012, + 0.005528958514332771, + 0.008655152283608913, + 0.008484390564262867, + 0.008403475396335125, + 0.004655145108699799, + 0.013335577212274075, + 0.008968818932771683, + 0.0020404676906764507, + 0.0016333576058968902, + 0.0022767321206629276, + 0.0018942169845104218, + 0.0023671204689890146, + 0.0017172590596601367, + 0.0028737173415720463, + 0.002102409955114126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.1428947001695633 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "still, dramatic overhead close-up, captured in stunning 8K resolution and extreme slow motion, like a decisive moment from a World Cup match, focuses on the surface of fertile, dark, moist soil, slowly rising and cracking under a force rising from beneath the surface. Suddenly, with an unstoppable, primal force, a vibrant, emerald green sprout bursts through the earth, its two cotyledons curled tightly like clenched fists, ready to announce its arrival to the world. With its emergence, countless individual soil particles, from fine dust to tiny clods of earth, are violently hurled upward and outward, forming a spectacular, solidified cloud of explosions in the air. The rough texture and irregular shape of each particle are clearly visible in the ultra-high resolution. A strong beam of early morning sunlight slants into the frame, outlining the flying soil particles with golden hues and making the shimmering dewdrops covering the sprout sparkle like tiny diamonds, contrasting sharply against the deep earthy background. The entire scene is filled with a primitive sense of power and speed, transforming a microscopic process of life birth into a grand, tension-filled and shocking visual epic.", + "seed": 30337, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_6/index_7679_seed_30337_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.5496033245697618, + "vlm_logprob": -0.5496033245697618, + "vlm_yes_no_margin": 2.699999952316284, + "vlm_token_credit": -0.5496033245697618, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015054537914693356 + ], + "opposite_logprob_tokens": [ + -5.265054702758789 + ], + "logprob_reward": -0.015054537914693356, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.015054537914693356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08445166051387787, + 0.1915956288576126, + 0.09085895121097565, + 0.043287597596645355, + 0.14846736192703247, + 0.04810371622443199, + 0.04419542849063873, + 0.2053872048854828, + 0.09763716906309128, + 0.04601525142788887 + ], + "credit_map": [ + 0.011336829513311386, + 0.011304454877972603, + 0.010018211789429188, + 0.009903302416205406, + 0.012193689122796059, + 0.009311084635555744, + 0.008289684541523457, + 0.01209440641105175, + 0.02146332897245884, + 0.026539461687207222, + 0.04041583836078644, + 0.012218432500958443, + 0.011675424873828888, + 0.04899813234806061, + 0.01206283736974001, + 0.018222162500023842, + 0.024035442620515823, + 0.009876709431409836, + 0.019628485664725304, + 0.007343962322920561, + 0.0066223954781889915, + 0.010124544613063335, + 0.007972952909767628, + 0.005254458636045456, + 0.0043229334987699986, + 0.005037966184318066, + 0.007258166093379259, + 0.0032281847670674324, + 0.004460812080651522, + 0.007249782793223858, + 0.00679559214040637, + 0.004934162832796574, + 0.015934370458126068, + 0.02068375237286091, + 0.019123075529932976, + 0.011641325429081917, + 0.01756400428712368, + 0.023916861042380333, + 0.016671570017933846, + 0.02293241210281849, + 0.004301984794437885, + 0.01150103472173214, + 0.004926623310893774, + 0.003288471605628729, + 0.004695952869951725, + 0.008104801177978516, + 0.003173019038513303, + 0.00811182800680399, + 0.0061393058858811855, + 0.003836167510598898, + 0.006087149493396282, + 0.005278751254081726, + 0.004496071022003889, + 0.004803522489964962, + 0.007456873543560505, + 0.00609758822247386, + 0.05423102155327797, + 0.024625297635793686, + 0.03470771014690399, + 0.022252274677157402, + 0.008598103187978268, + 0.02564823441207409, + 0.008248884230852127, + 0.02707567997276783, + 0.016082918271422386, + 0.021061040461063385, + 0.010830686427652836, + 0.009779877960681915, + 0.012245312333106995, + 0.008466629311442375, + 0.010872023180127144, + 0.008298678323626518, + 0.01112352591007948, + 0.009304225444793701, + 0.004490935709327459, + 0.003511236747726798, + 0.0038854002486914396, + 0.007394910790026188, + 0.0029045860283076763, + 0.0034004314802587032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.015054537914693356 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008669713512063026 + ], + "opposite_logprob_tokens": [ + -4.758669853210449 + ], + "logprob_reward": -0.008669713512063026, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008669713512063026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15103736519813538, + 0.16347308456897736, + 0.13239002227783203, + 0.062195274978876114, + 0.14152908325195312, + 0.07655143737792969, + 0.06115354970097542, + 0.06440882384777069, + 0.08028127253055573, + 0.06698001176118851 + ], + "credit_map": [ + 0.014291941188275814, + 0.024968313053250313, + 0.022758252918720245, + 0.022089719772338867, + 0.020037135109305382, + 0.013803270645439625, + 0.011007540859282017, + 0.022081218659877777, + 0.01931166648864746, + 0.023333914577960968, + 0.032393891364336014, + 0.010418633930385113, + 0.009929104708135128, + 0.03935817629098892, + 0.010523146949708462, + 0.018204567953944206, + 0.03382839635014534, + 0.011347784660756588, + 0.0254351943731308, + 0.023879457265138626, + 0.006801854353398085, + 0.013862268067896366, + 0.00699315732344985, + 0.010241924785077572, + 0.008052576333284378, + 0.008224346674978733, + 0.007754836231470108, + 0.005697716027498245, + 0.005511724390089512, + 0.009529944509267807, + 0.010956228710711002, + 0.006467909086495638, + 0.021202554926276207, + 0.020947979763150215, + 0.019966065883636475, + 0.012085494585335255, + 0.01792929135262966, + 0.01905715838074684, + 0.017712410539388657, + 0.012628142721951008, + 0.0070121935568749905, + 0.014328008517622948, + 0.009106994606554508, + 0.008940205909311771, + 0.005735184531658888, + 0.012843496166169643, + 0.008520923554897308, + 0.010064437985420227, + 0.00711882533505559, + 0.0062377662397921085, + 0.008287940174341202, + 0.006999066565185785, + 0.006555001717060804, + 0.008389120921492577, + 0.00949610024690628, + 0.008069733157753944, + 0.017499936744570732, + 0.007740041706711054, + 0.00889617670327425, + 0.006592192687094212, + 0.003919803071767092, + 0.00821937806904316, + 0.0037328095640987158, + 0.00780849251896143, + 0.01336126308888197, + 0.009998869150876999, + 0.009750046767294407, + 0.010097797960042953, + 0.008275368250906467, + 0.009705320000648499, + 0.010870397090911865, + 0.008222215808928013, + 0.011969603598117828, + 0.006929137744009495, + 0.011556054465472698, + 0.007186429109424353, + 0.0056224362924695015, + 0.009044560603797436, + 0.005021671298891306, + 0.009650127962231636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.008669713512063026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018204394727945328 + ], + "opposite_logprob_tokens": [ + -4.018204212188721 + ], + "logprob_reward": -0.018204394727945328, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018204394727945328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11313562095165253, + 0.0662224069237709, + 0.1766505390405655, + 0.06617570668458939, + 0.20520246028900146, + 0.05475828796625137, + 0.03115043416619301, + 0.0872301310300827, + 0.16290606558322906, + 0.03656835854053497 + ], + "credit_map": [ + 0.009845895692706108, + 0.011445085518062115, + 0.012759710662066936, + 0.019484812393784523, + 0.014692396856844425, + 0.009744162671267986, + 0.016214706003665924, + 0.018948860466480255, + 0.006838607136160135, + 0.0062450384721159935, + 0.012558718211948872, + 0.006125766318291426, + 0.007817186415195465, + 0.012834812514483929, + 0.006719870027154684, + 0.007082407828420401, + 0.03496571630239487, + 0.01924828253686428, + 0.02472233958542347, + 0.045627426356077194, + 0.011324388906359673, + 0.01469457894563675, + 0.013870843686163425, + 0.012196962721645832, + 0.007989484816789627, + 0.008299492299556732, + 0.011525464244186878, + 0.004903819411993027, + 0.006333786528557539, + 0.011937241069972515, + 0.007503692060709, + 0.007682729046791792, + 0.017212258651852608, + 0.014707091264426708, + 0.02363639883697033, + 0.0315103754401207, + 0.010657276026904583, + 0.03532588481903076, + 0.018606996163725853, + 0.05354617163538933, + 0.004540910013020039, + 0.013545414432883263, + 0.0056353784166276455, + 0.0043096840381622314, + 0.006329053547233343, + 0.005823762156069279, + 0.003853355534374714, + 0.010720728896558285, + 0.0034107984974980354, + 0.0040269093587994576, + 0.0031927579548209906, + 0.003908572252839804, + 0.004590088035911322, + 0.0030657784081995487, + 0.0059571582823991776, + 0.0029983704444020987, + 0.02149573713541031, + 0.01089286059141159, + 0.012983890250325203, + 0.010110250674188137, + 0.005722593516111374, + 0.00882311724126339, + 0.005565432831645012, + 0.01163624506443739, + 0.024446405470371246, + 0.03173912316560745, + 0.01887611486017704, + 0.016121957451105118, + 0.021195752546191216, + 0.018176674842834473, + 0.016265593469142914, + 0.01608443073928356, + 0.004002116154879332, + 0.004760802257806063, + 0.004392177332192659, + 0.00420794915407896, + 0.004977331031113863, + 0.00482534384354949, + 0.005444864742457867, + 0.003957773093134165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "selected_reward": -0.018204394727945328 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.5789475440979004 + ], + "opposite_logprob_tokens": [ + -0.07894758135080338 + ], + "logprob_reward": -2.5789475440979004, + "yes_no_margin_reward": -2.5, + "token_credit_reward": -2.5789475440979004, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07586321979761124, + 0.3448379337787628, + 0.16216668486595154, + 0.05829031392931938, + 0.06863358616828918, + 0.041731301695108414, + 0.03594515845179558, + 0.03737654164433479, + 0.10545257478952408, + 0.0697026252746582 + ], + "credit_map": [ + 0.00877743773162365, + 0.010249546729028225, + 0.008844569325447083, + 0.011391034349799156, + 0.009296109899878502, + 0.007414190098643303, + 0.011141984723508358, + 0.008748345077037811, + 0.0427282489836216, + 0.056373339146375656, + 0.042514849454164505, + 0.028014780953526497, + 0.033049073070287704, + 0.06601767987012863, + 0.026694228872656822, + 0.0494457446038723, + 0.015976689755916595, + 0.0257756095379591, + 0.023677080869674683, + 0.03202296420931816, + 0.014396805316209793, + 0.03076695278286934, + 0.01117886882275343, + 0.008371712639927864, + 0.009133908897638321, + 0.006909584626555443, + 0.006463488098233938, + 0.006267570424824953, + 0.0057767340913414955, + 0.011445166543126106, + 0.007525273598730564, + 0.004768585320562124, + 0.006715089548379183, + 0.007464995142072439, + 0.008527526631951332, + 0.015718406066298485, + 0.0058151185512542725, + 0.007268785499036312, + 0.008854899555444717, + 0.00826877262443304, + 0.004448858089745045, + 0.0046621947549283504, + 0.004989605396986008, + 0.005676389206200838, + 0.004997385665774345, + 0.006298114079982042, + 0.00516253337264061, + 0.005496223457157612, + 0.0038476574700325727, + 0.002370345639064908, + 0.00415503466501832, + 0.0048186276108026505, + 0.00419836537912488, + 0.003657326567918062, + 0.0061279479414224625, + 0.006769854109734297, + 0.007920531556010246, + 0.005466712638735771, + 0.004350787494331598, + 0.003402149537578225, + 0.0045381817035377026, + 0.003236559219658375, + 0.0040386756882071495, + 0.004422941245138645, + 0.01527001615613699, + 0.011661979369819164, + 0.014076558873057365, + 0.01097555086016655, + 0.014614438638091087, + 0.012095418758690357, + 0.010782350786030293, + 0.015976274386048317, + 0.013679390773177147, + 0.012974239885807037, + 0.007450577337294817, + 0.006388416979461908, + 0.009805838577449322, + 0.007812208030372858, + 0.00576404482126236, + 0.0058279093354940414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -2.5789475440979004 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle an overhead close-up shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soil surface dark, moist, and fertile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant emerald green sprout with two curled cotyledons burst through the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are soil particles violently hurled upward forming a solidified cloud of explosions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a strong beam of early morning sunlight slanting into the frame to outline particles and dewdrops?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12714043259620667 + ], + "opposite_logprob_tokens": [ + -2.127140522003174 + ], + "logprob_reward": -0.12714043259620667, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12714043259620667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1427454799413681, + 0.24358795583248138, + 0.12062177807092667, + 0.0317394882440567, + 0.11179570108652115, + 0.04442237317562103, + 0.031490784138441086, + 0.09642952680587769, + 0.056847020983695984, + 0.1203199028968811 + ], + "credit_map": [ + 0.016236625611782074, + 0.018199894577264786, + 0.018834326416254044, + 0.015666671097278595, + 0.021609291434288025, + 0.014262494631111622, + 0.014414548873901367, + 0.023521631956100464, + 0.02546856366097927, + 0.03670084476470947, + 0.04116067662835121, + 0.02653631381690502, + 0.027628904208540916, + 0.03622738644480705, + 0.0175357386469841, + 0.03232952207326889, + 0.025051329284906387, + 0.010402906686067581, + 0.026822775602340698, + 0.009643248282372952, + 0.011567650362849236, + 0.019817650318145752, + 0.010695360600948334, + 0.00662085460498929, + 0.0037120000924915075, + 0.005143322516232729, + 0.004249005578458309, + 0.0026295986026525497, + 0.004044781904667616, + 0.003431714838370681, + 0.004641808569431305, + 0.0038872542791068554, + 0.01609226129949093, + 0.015120640397071838, + 0.014701833948493004, + 0.009838507510721684, + 0.010715240612626076, + 0.014559674076735973, + 0.011141740716993809, + 0.01962580904364586, + 0.003681971225887537, + 0.011148431338369846, + 0.003776376135647297, + 0.0036573184188455343, + 0.003742654575034976, + 0.005615996662527323, + 0.0032592182978987694, + 0.009540405124425888, + 0.003302445402368903, + 0.0017806069226935506, + 0.0035754044074565172, + 0.003596771042793989, + 0.0031935942824929953, + 0.0030258374754339457, + 0.006968896836042404, + 0.006047229748219252, + 0.02522153966128826, + 0.010682784952223301, + 0.015173681080341339, + 0.010154460556805134, + 0.006302272900938988, + 0.010566114448010921, + 0.0057699959725141525, + 0.012558681890368462, + 0.00904762465506792, + 0.00595055939629674, + 0.0068694897927343845, + 0.00540220458060503, + 0.006368973758071661, + 0.006920407991856337, + 0.0063397628255188465, + 0.009947996586561203, + 0.022488052025437355, + 0.029095180332660675, + 0.013443592004477978, + 0.009381404146552086, + 0.011629815213382244, + 0.014591177925467491, + 0.009394641034305096, + 0.010296053253114223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.12714043259620667 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_7/index_tensor([8789, 8789])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_7/index_tensor([8789, 8789])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4658de557f60656c96f3fffde5e9dc1312112804 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_7/index_tensor([8789, 8789])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows a female warrior as she runs from left to right, advancing across the vast battlefield with a powerful yet somewhat weighted demeanor. Her long, dark brown hair, braided into intricate braids, whips violently as she runs. Her face is stern, stained with dust and sweat. Her eyes gaze sharply ahead, her lips pursed, revealing an unyielding resolve. She clads in a suit of ancient, battle-worn armor, a blend of bronze and leather. The breastplate is faintly engraved with totems, and the shoulder pads and wrist guards are scarred and dented. Beneath the armor lies a crimson, coarse cloth kilt, its hems billowing with her movements. Her arms swing vigorously for balance, each step sinking deeply into the muddy, uneven ground beneath her feet, sending small splashes of mud flying. In the background, a dark, smoke-filled sky rises. Sporadic fires burn in the distance, and broken flags and scattered weapons litter the desolate battlefield. The entire image exhibits a distinct sense of retro film grain, with tones tending towards low-saturation sepia and gray, colors that seem faded through the years, and slight vignetting and irregular light leaks at the edges, which together create a strong sense of epic and tragic atmosphere.", + "seed": 4146, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_7/index_8789_seed_4146_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03727170417550951, + "vlm_logprob": -0.03727170417550951, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.03727170417550951, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11955635994672775 + ], + "opposite_logprob_tokens": [ + -2.244556427001953 + ], + "logprob_reward": -0.11955635994672775, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11955635994672775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06762290000915527, + 0.02513735555112362, + 0.040131404995918274, + 0.5560741424560547, + 0.04550773277878761, + 0.1059158444404602, + 0.03166574612259865, + 0.03232986107468605, + 0.07793569564819336, + 0.017679400742053986 + ], + "credit_map": [ + 0.005065161269158125, + 0.009065475314855576, + 0.007677730638533831, + 0.009259471669793129, + 0.009079056791961193, + 0.011940723285079002, + 0.00876542180776596, + 0.006769861094653606, + 0.0022015131544321775, + 0.0029866963159292936, + 0.0019259199034422636, + 0.0020638681016862392, + 0.006633727811276913, + 0.0038494374603033066, + 0.0033109711948782206, + 0.0021652202121913433, + 0.003791158553212881, + 0.0032687047496438026, + 0.004486434161663055, + 0.003066661302000284, + 0.005305181257426739, + 0.0048316302709281445, + 0.008984345942735672, + 0.006397288292646408, + 0.06378290802240372, + 0.027956601232290268, + 0.1482587307691574, + 0.036472804844379425, + 0.04041182994842529, + 0.06767221540212631, + 0.041682202368974686, + 0.1298367828130722, + 0.005100256763398647, + 0.005171184428036213, + 0.005097212735563517, + 0.005184919573366642, + 0.008276212960481644, + 0.005265499465167522, + 0.00585007993504405, + 0.0055623650550842285, + 0.007574467454105616, + 0.005642176140099764, + 0.03673524037003517, + 0.003909260034561157, + 0.005575761664658785, + 0.038691338151693344, + 0.0038666396867483854, + 0.00392095698043704, + 0.0029578928370028734, + 0.002852214965969324, + 0.00994928926229477, + 0.002812444930896163, + 0.00253066117875278, + 0.0025895442813634872, + 0.0019320885185152292, + 0.006041606422513723, + 0.0022590141743421555, + 0.003563906066119671, + 0.006630733143538237, + 0.004305476788431406, + 0.005422916263341904, + 0.003717944724485278, + 0.003381201531738043, + 0.0030486639589071274, + 0.009045671671628952, + 0.00997155625373125, + 0.026989366859197617, + 0.005774631164968014, + 0.004215300548821688, + 0.013876991346478462, + 0.003342842683196068, + 0.004719334654510021, + 0.0019087966065853834, + 0.0017697998555377126, + 0.001995128346607089, + 0.0018586545484140515, + 0.002789880381897092, + 0.0019223256967961788, + 0.002748944563791156, + 0.0026858700439333916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.11955635994672775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016084907576441765 + ], + "opposite_logprob_tokens": [ + -4.141084671020508 + ], + "logprob_reward": -0.016084907576441765, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016084907576441765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1788983792066574, + 0.041208893060684204, + 0.05450912192463875, + 0.2789742350578308, + 0.06934194266796112, + 0.1345319151878357, + 0.06402558088302612, + 0.04259239137172699, + 0.10341279953718185, + 0.032504670321941376 + ], + "credit_map": [ + 0.01636386290192604, + 0.02262442745268345, + 0.01726401224732399, + 0.01701582409441471, + 0.0312880240380764, + 0.04307162016630173, + 0.016420407220721245, + 0.014850219711661339, + 0.004626201931387186, + 0.008127624168992043, + 0.004169074352830648, + 0.006379382219165564, + 0.005216267891228199, + 0.004368848633021116, + 0.004345771856606007, + 0.0039757248014211655, + 0.0050104218535125256, + 0.005907839629799128, + 0.006296643055975437, + 0.0065039959736168385, + 0.005238005891442299, + 0.0067551350221037865, + 0.009314501658082008, + 0.009482585825026035, + 0.03141162171959877, + 0.013632340356707573, + 0.07210250198841095, + 0.021329015493392944, + 0.01915290579199791, + 0.03266475722193718, + 0.024598732590675354, + 0.06408239901065826, + 0.007733931764960289, + 0.007961353287100792, + 0.006589911878108978, + 0.012714252807199955, + 0.008525214157998562, + 0.009095611982047558, + 0.009294521994888783, + 0.007427154574543238, + 0.009968362748622894, + 0.006342356558889151, + 0.048907388001680374, + 0.0059010605327785015, + 0.006664615590125322, + 0.04605250433087349, + 0.005515444092452526, + 0.005180189851671457, + 0.0071029057726264, + 0.009494048543274403, + 0.007925929501652718, + 0.005870627239346504, + 0.006271094083786011, + 0.006779203657060862, + 0.012143805623054504, + 0.00843797530978918, + 0.00597127340734005, + 0.004638270009309053, + 0.005844839382916689, + 0.005467553623020649, + 0.006394167430698872, + 0.00536838173866272, + 0.005197873339056969, + 0.0037100359331816435, + 0.013150709681212902, + 0.008732279762625694, + 0.03895791620016098, + 0.0073326630517840385, + 0.004569706041365862, + 0.021648991852998734, + 0.004346039611846209, + 0.004674504045397043, + 0.004681553225964308, + 0.004400057252496481, + 0.004967271815985441, + 0.0034353267401456833, + 0.004664287436753511, + 0.0049466523341834545, + 0.003092310857027769, + 0.002317216480150819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.016084907576441765 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306042939424515 + ], + "opposite_logprob_tokens": [ + -3.168060541152954 + ], + "logprob_reward": -0.04306042939424515, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306042939424515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08964407444000244, + 0.02240813709795475, + 0.06973300129175186, + 0.5257568359375, + 0.016996467486023903, + 0.026646126061677933, + 0.034417666494846344, + 0.05502406507730484, + 0.12416720390319824, + 0.0352063849568367 + ], + "credit_map": [ + 0.008648701012134552, + 0.01280012633651495, + 0.01392568089067936, + 0.00856892392039299, + 0.014681430533528328, + 0.011778376996517181, + 0.008845151402056217, + 0.01039568055421114, + 0.002013009274378419, + 0.0037188928108662367, + 0.002329868031665683, + 0.002570497803390026, + 0.004876923281699419, + 0.002431518631055951, + 0.0026771845296025276, + 0.0017902422696352005, + 0.007111174985766411, + 0.009594479575753212, + 0.010408162139356136, + 0.009073402732610703, + 0.00563250994309783, + 0.00926225632429123, + 0.009299032390117645, + 0.009351987391710281, + 0.05833832547068596, + 0.024461451917886734, + 0.1439252644777298, + 0.03926019370555878, + 0.03491467982530594, + 0.06351616978645325, + 0.04607667028903961, + 0.11526411771774292, + 0.002209929283708334, + 0.0020113810896873474, + 0.0022109663113951683, + 0.0022667201701551676, + 0.0024497273843735456, + 0.0021109874360263348, + 0.00223886058665812, + 0.0014978941762819886, + 0.0023864165414124727, + 0.001819438417442143, + 0.007596626877784729, + 0.001479733269661665, + 0.001566514722071588, + 0.008540241047739983, + 0.001714428886771202, + 0.0015427242033183575, + 0.002680369885638356, + 0.002596269128844142, + 0.010946722701191902, + 0.002263392321765423, + 0.0017346427775919437, + 0.002493947045877576, + 0.00406072149053216, + 0.00764160230755806, + 0.004350628703832626, + 0.005630665924400091, + 0.010835709981620312, + 0.009623250924050808, + 0.00822541955858469, + 0.006833415944129229, + 0.004931014031171799, + 0.004593959543853998, + 0.011698245070874691, + 0.008132195100188255, + 0.044901493936777115, + 0.008335578255355358, + 0.0054523092694580555, + 0.03426764905452728, + 0.004935345146805048, + 0.0064443848095834255, + 0.0032593708019703627, + 0.0034235725179314613, + 0.004213492386043072, + 0.004759719595313072, + 0.0054520801641047, + 0.003855459624901414, + 0.005919938907027245, + 0.004322751425206661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.04306042939424515 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001709786825813353 + ], + "opposite_logprob_tokens": [ + -6.376709938049316 + ], + "logprob_reward": -0.001709786825813353, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001709786825813353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10605882108211517, + 0.022987129166722298, + 0.013146638870239258, + 0.2785473167896271, + 0.027072656899690628, + 0.3225415349006653, + 0.0668557807803154, + 0.038751617074012756, + 0.06749875098466873, + 0.05653972923755646 + ], + "credit_map": [ + 0.010481052100658417, + 0.009781748987734318, + 0.017124611884355545, + 0.010399573482573032, + 0.016816861927509308, + 0.018373113125562668, + 0.012258125469088554, + 0.010823724791407585, + 0.0023967840243130922, + 0.002303148154169321, + 0.003170957788825035, + 0.0018186521483585238, + 0.004059786908328533, + 0.0043655820190906525, + 0.0027381260879337788, + 0.002134093549102545, + 0.0017799464985728264, + 0.0015964164631441236, + 0.001292901812121272, + 0.0025282998103648424, + 0.0012804112629964948, + 0.0019124102545902133, + 0.0016159240622073412, + 0.0011403284734115005, + 0.03243197873234749, + 0.013584206812083721, + 0.07738526910543442, + 0.017396096140146255, + 0.021564727649092674, + 0.035085853189229965, + 0.015486519783735275, + 0.06561265885829926, + 0.003455100581049919, + 0.0031457694713026285, + 0.004443801008164883, + 0.0032017838675528765, + 0.004546979907900095, + 0.0030681623611599207, + 0.003127409378066659, + 0.0020836512558162212, + 0.021953221410512924, + 0.01513922680169344, + 0.11927852034568787, + 0.012918217107653618, + 0.014245939441025257, + 0.11690043658018112, + 0.011941774748265743, + 0.010164217092096806, + 0.006876900792121887, + 0.0070228311233222485, + 0.02422877587378025, + 0.004776178393512964, + 0.0023797559551894665, + 0.00423450255766511, + 0.003468588925898075, + 0.01386825181543827, + 0.004556896630674601, + 0.005133995786309242, + 0.0073919836431741714, + 0.006973229814320803, + 0.0038270042277872562, + 0.004699887707829475, + 0.0027994425036013126, + 0.0033691765274852514, + 0.008569625206291676, + 0.005486364010721445, + 0.024689197540283203, + 0.004795378539711237, + 0.0033116189297288656, + 0.014306321740150452, + 0.002826235257089138, + 0.003514006966724992, + 0.007475048303604126, + 0.008700713515281677, + 0.007604760117828846, + 0.006238247733563185, + 0.007520387414842844, + 0.008292115293443203, + 0.005383952055126429, + 0.0053245071321725845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.001709786825813353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947037134319544 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947037134319544, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947037134319544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09980237483978271, + 0.02973199635744095, + 0.06345892697572708, + 0.054255902767181396, + 0.04619112238287926, + 0.30295929312705994, + 0.04480333998799324, + 0.06463954597711563, + 0.26451703906059265, + 0.029640382155776024 + ], + "credit_map": [ + 0.009132257662713528, + 0.0141415661200881, + 0.017295394092798233, + 0.007197501137852669, + 0.016190826892852783, + 0.015692327171564102, + 0.010285615921020508, + 0.009866894222795963, + 0.004402560647577047, + 0.003003216115757823, + 0.004514561500400305, + 0.0022205861750990152, + 0.0041211312636733055, + 0.0033905799500644207, + 0.003384404117241502, + 0.004694957751780748, + 0.006558703258633614, + 0.0061434064991772175, + 0.008138952776789665, + 0.004720783792436123, + 0.007744591683149338, + 0.007692641578614712, + 0.012068602256476879, + 0.010391246527433395, + 0.005245145410299301, + 0.005142931826412678, + 0.0118044912815094, + 0.004219161346554756, + 0.005444594658911228, + 0.0064828358590602875, + 0.005139382090419531, + 0.010777361690998077, + 0.007605349645018578, + 0.004129921086132526, + 0.004774089902639389, + 0.0073044560849666595, + 0.005752693861722946, + 0.005540936719626188, + 0.005513850599527359, + 0.00556982820853591, + 0.02195785380899906, + 0.01421959325671196, + 0.10934488475322723, + 0.01236901618540287, + 0.014459185302257538, + 0.11048965156078339, + 0.011554423719644547, + 0.008564705029129982, + 0.0034960799384862185, + 0.004695937968790531, + 0.009678990580141544, + 0.0037862220779061317, + 0.0065397960133850574, + 0.0035398751497268677, + 0.00489375414326787, + 0.008172690868377686, + 0.009162357077002525, + 0.007937458343803883, + 0.010358784347772598, + 0.010668310336768627, + 0.007799644488841295, + 0.006998102180659771, + 0.006705745123326778, + 0.00500915152952075, + 0.03128964453935623, + 0.02311946079134941, + 0.09992242604494095, + 0.017530204728245735, + 0.010933038778603077, + 0.06188809499144554, + 0.008593752048909664, + 0.011240456253290176, + 0.003836365882307291, + 0.0038743868935853243, + 0.003526706248521805, + 0.0031264377757906914, + 0.00372499693185091, + 0.004092725925147533, + 0.00327264959923923, + 0.004186117090284824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.005947037134319544 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a female warrior as she runs from left to right, advancing across the vast battlefield with a powerful yet somewhat weighted demeanor. Her long, dark brown hair, braided into intricate braids, whips violently as she runs. Her face is stern, stained with dust and sweat. Her eyes gaze sharply ahead, her lips pursed, revealing an unyielding resolve. She clads in a suit of ancient, battle-worn armor, a blend of bronze and leather. The breastplate is faintly engraved with totems, and the shoulder pads and wrist guards are scarred and dented. Beneath the armor lies a crimson, coarse cloth kilt, its hems billowing with her movements. Her arms swing vigorously for balance, each step sinking deeply into the muddy, uneven ground beneath her feet, sending small splashes of mud flying. In the background, a dark, smoke-filled sky rises. Sporadic fires burn in the distance, and broken flags and scattered weapons litter the desolate battlefield. The entire image exhibits a distinct sense of retro film grain, with tones tending towards low-saturation sepia and gray, colors that seem faded through the years, and slight vignetting and irregular light leaks at the edges, which together create a strong sense of epic and tragic atmosphere.", + "seed": 4147, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_7/index_8789_seed_4147_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02353946310468018, + "vlm_logprob": -0.02353946310468018, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.02353946310468018, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07283422350883484 + ], + "opposite_logprob_tokens": [ + -2.6978342533111572 + ], + "logprob_reward": -0.07283422350883484, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07283422350883484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044227778911590576, + 0.014321191236376762, + 0.011387187987565994, + 0.03845980390906334, + 0.021368559449911118, + 0.011366399005055428, + 0.17648445069789886, + 0.04862813279032707, + 0.6174095869064331, + 0.016346905380487442 + ], + "credit_map": [ + 0.0037177943158894777, + 0.004724615719169378, + 0.005790552590042353, + 0.004127137828618288, + 0.005620215088129044, + 0.00483337789773941, + 0.009365573525428772, + 0.006048512179404497, + 0.0019525201059877872, + 0.0014621054287999868, + 0.003195397322997451, + 0.0013855218421667814, + 0.001604805700480938, + 0.0017476220382377505, + 0.0016406517243012786, + 0.0013325671898201108, + 0.001224192907102406, + 0.0014522609999403358, + 0.002181250136345625, + 0.0011857140343636274, + 0.0012433272786438465, + 0.0013137481873854995, + 0.0017803115770220757, + 0.001006383216008544, + 0.002728249877691269, + 0.003913501743227243, + 0.0056465319357812405, + 0.004870137665420771, + 0.007617794908583164, + 0.004069135524332523, + 0.004614727105945349, + 0.004999724682420492, + 0.001350054400973022, + 0.002308045979589224, + 0.002182932570576668, + 0.0021228049881756306, + 0.001888450002297759, + 0.004254487343132496, + 0.00322489975951612, + 0.004036883823573589, + 0.0011026496067643166, + 0.0012194972950965166, + 0.0024381664115935564, + 0.0009350057807751, + 0.0019045770168304443, + 0.001331899780780077, + 0.0010918608168140054, + 0.0013427423546090722, + 0.034721653908491135, + 0.011728017590939999, + 0.02907395549118519, + 0.011679751798510551, + 0.013022979721426964, + 0.022530844435095787, + 0.014308363199234009, + 0.03941889479756355, + 0.004393754992634058, + 0.004135014023631811, + 0.005428022239357233, + 0.004521792754530907, + 0.0076721650548279285, + 0.006175273563712835, + 0.0058768196031451225, + 0.010425292886793613, + 0.07688239216804504, + 0.03195418044924736, + 0.13355065882205963, + 0.03735857084393501, + 0.054511841386556625, + 0.07708095014095306, + 0.07285445183515549, + 0.13321655988693237, + 0.0021529204677790403, + 0.002345045330002904, + 0.002676085103303194, + 0.00124449806753546, + 0.0023756632581353188, + 0.002395659452304244, + 0.001971933990716934, + 0.0011851004092022777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.07283422350883484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549715116620064 + ], + "opposite_logprob_tokens": [ + -4.387549877166748 + ], + "logprob_reward": -0.012549715116620064, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549715116620064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11457476764917374, + 0.0423034131526947, + 0.0767071396112442, + 0.051963191479444504, + 0.033181700855493546, + 0.03354901075363159, + 0.19443881511688232, + 0.08860815316438675, + 0.2866039574146271, + 0.0780697762966156 + ], + "credit_map": [ + 0.012958001345396042, + 0.008618457242846489, + 0.017476236447691917, + 0.011666971258819103, + 0.011207959614694118, + 0.014520646072924137, + 0.026753341779112816, + 0.011373157612979412, + 0.004689101595431566, + 0.004757757764309645, + 0.007009900175035, + 0.005086828954517841, + 0.005432680249214172, + 0.005426669958978891, + 0.005538123659789562, + 0.004362354055047035, + 0.004156017675995827, + 0.005977629218250513, + 0.021459326148033142, + 0.0064065344631671906, + 0.0046791378408670425, + 0.009494923055171967, + 0.01464084256440401, + 0.009892729111015797, + 0.004158042836934328, + 0.0052124024368822575, + 0.006165781058371067, + 0.008218598552048206, + 0.0067765479907393456, + 0.005914690904319286, + 0.007937588728964329, + 0.007579541299492121, + 0.0024372227489948273, + 0.004415243864059448, + 0.004835178144276142, + 0.004711652174592018, + 0.003247938584536314, + 0.004306375049054623, + 0.004474314395338297, + 0.004753777291625738, + 0.0030432601924985647, + 0.003672461723908782, + 0.005143200047314167, + 0.004710599314421415, + 0.005097154062241316, + 0.003680552588775754, + 0.004949474707245827, + 0.003252306953072548, + 0.027677619829773903, + 0.023279892280697823, + 0.02847948856651783, + 0.018446315079927444, + 0.02016196772456169, + 0.01982973888516426, + 0.02403344213962555, + 0.032530345022678375, + 0.008158494718372822, + 0.01044962927699089, + 0.010839761234819889, + 0.009690393693745136, + 0.012244729325175285, + 0.010407806374132633, + 0.015329830348491669, + 0.011487508192658424, + 0.03603251278400421, + 0.014906704425811768, + 0.061345070600509644, + 0.01840517483651638, + 0.023305611684918404, + 0.03606046736240387, + 0.03355195000767708, + 0.06299646198749542, + 0.008883442729711533, + 0.010828249156475067, + 0.009813717566430569, + 0.008513685315847397, + 0.010567592456936836, + 0.01048327051103115, + 0.012572060339152813, + 0.0064077600836753845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.012549715116620064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346080005168915 + ], + "opposite_logprob_tokens": [ + -3.65134596824646 + ], + "logprob_reward": -0.026346080005168915, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346080005168915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23589318990707397, + 0.042736321687698364, + 0.04987768828868866, + 0.07330604642629623, + 0.07185524702072144, + 0.038785047829151154, + 0.14244696497917175, + 0.037967145442962646, + 0.2650097906589508, + 0.042122531682252884 + ], + "credit_map": [ + 0.02277330681681633, + 0.02478504553437233, + 0.029941989108920097, + 0.02381647378206253, + 0.027739912271499634, + 0.030868614092469215, + 0.04729674383997917, + 0.028671113774180412, + 0.00658242916688323, + 0.00457541411742568, + 0.008091314695775509, + 0.0038106171414256096, + 0.004633219446986914, + 0.005987517070025206, + 0.004760006442666054, + 0.004295803140848875, + 0.004311776719987392, + 0.004242200404405594, + 0.013598957099020481, + 0.00429381662979722, + 0.004427329637110233, + 0.005699774716049433, + 0.006420021876692772, + 0.006883809808641672, + 0.007276431191712618, + 0.009051001630723476, + 0.009800259955227375, + 0.00855256337672472, + 0.01167380716651678, + 0.009044565260410309, + 0.007961767725646496, + 0.009945650584995747, + 0.005043704994022846, + 0.005149581469595432, + 0.006017484702169895, + 0.005805138498544693, + 0.005769438110291958, + 0.018630189821124077, + 0.006665864028036594, + 0.018773850053548813, + 0.002471336629241705, + 0.0034103195648640394, + 0.00554380239918828, + 0.0032168147154152393, + 0.004176159389317036, + 0.009026438929140568, + 0.003568715648725629, + 0.007371458224952221, + 0.02847912162542343, + 0.011728380806744099, + 0.023832334205508232, + 0.009541494771838188, + 0.011505579575896263, + 0.017157191410660744, + 0.010446956381201744, + 0.029755916446447372, + 0.0038771070539951324, + 0.0047989049926400185, + 0.004897051490843296, + 0.003784986212849617, + 0.006546685937792063, + 0.004429701715707779, + 0.00410106685012579, + 0.005531637463718653, + 0.02851344645023346, + 0.01646575704216957, + 0.05975327640771866, + 0.022295141592621803, + 0.02175881527364254, + 0.03546791523694992, + 0.03608443960547447, + 0.044670991599559784, + 0.006209790240973234, + 0.004407807253301144, + 0.007744962349534035, + 0.0032108176965266466, + 0.005103548057377338, + 0.005897653289139271, + 0.005005809012800455, + 0.004542145412415266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.026346080005168915 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013325868640094995 + ], + "opposite_logprob_tokens": [ + -6.626332759857178 + ], + "logprob_reward": -0.0013325868640094995, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013325868640094995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16778378188610077, + 0.09217792749404907, + 0.18132784962654114, + 0.12948454916477203, + 0.1232748031616211, + 0.08773451298475266, + 0.06909211724996567, + 0.04581863433122635, + 0.06787429749965668, + 0.03543149307370186 + ], + "credit_map": [ + 0.029108623042702675, + 0.015326873399317265, + 0.03390054777264595, + 0.012043788097798824, + 0.012324315495789051, + 0.027548445388674736, + 0.020650537684559822, + 0.016880659386515617, + 0.011965515092015266, + 0.013296913355588913, + 0.018031319603323936, + 0.00882738083600998, + 0.009129333309829235, + 0.012792583554983139, + 0.007541177328675985, + 0.010593706741929054, + 0.01720360666513443, + 0.017018239945173264, + 0.05186464637517929, + 0.015155624598264694, + 0.01048764493316412, + 0.021693017333745956, + 0.018634002655744553, + 0.029271064326167107, + 0.016819441691040993, + 0.01636572740972042, + 0.022966761142015457, + 0.015791088342666626, + 0.01698431745171547, + 0.015213965438306332, + 0.010092026554048061, + 0.015251216478645802, + 0.008739231154322624, + 0.009063608013093472, + 0.012193769216537476, + 0.008678137324750423, + 0.011090236715972424, + 0.03345721587538719, + 0.007258571684360504, + 0.032794032245874405, + 0.006342833861708641, + 0.003980529494583607, + 0.010908698663115501, + 0.0038566086441278458, + 0.004522376228123903, + 0.027125611901283264, + 0.0044723316095769405, + 0.026525523513555527, + 0.012122243642807007, + 0.008010398596525192, + 0.010589645244181156, + 0.005399127956479788, + 0.00754421204328537, + 0.00680801784619689, + 0.007155098952353001, + 0.011463379487395287, + 0.003686073236167431, + 0.0035395605955272913, + 0.009367715567350388, + 0.003439431544393301, + 0.004751097410917282, + 0.004831143654882908, + 0.0071647800505161285, + 0.009038833901286125, + 0.010230431333184242, + 0.003963455092161894, + 0.008622194640338421, + 0.0036002458073198795, + 0.008319961838424206, + 0.008588175289332867, + 0.008329068310558796, + 0.01622076705098152, + 0.004226457327604294, + 0.003634953871369362, + 0.0045954142697155476, + 0.0036813754122704268, + 0.004805588163435459, + 0.004772638902068138, + 0.0052533275447785854, + 0.00446173595264554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0013325868640094995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634710028767586 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004634710028767586, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634710028767586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1036761924624443, + 0.04281798005104065, + 0.046907734125852585, + 0.16724270582199097, + 0.09009964019060135, + 0.0527506060898304, + 0.10476788133382797, + 0.07725320011377335, + 0.22093456983566284, + 0.09354947507381439 + ], + "credit_map": [ + 0.01193323265761137, + 0.013067325577139854, + 0.012643026188015938, + 0.012895746156573296, + 0.012642545625567436, + 0.012378825806081295, + 0.014913447201251984, + 0.013202046044170856, + 0.006059224251657724, + 0.0058512999676167965, + 0.007558496203273535, + 0.004407161381095648, + 0.005498387850821018, + 0.005067133344709873, + 0.004527132958173752, + 0.0038491415325552225, + 0.006641772575676441, + 0.0052725775167346, + 0.011633021757006645, + 0.0036049671471118927, + 0.0032685045152902603, + 0.0052915518172085285, + 0.0061470926739275455, + 0.005048245191574097, + 0.014034304767847061, + 0.021313762292265892, + 0.03529195860028267, + 0.022124752402305603, + 0.022514330223202705, + 0.01887127012014389, + 0.011812733486294746, + 0.021279599517583847, + 0.0066259209997951984, + 0.0064112176187336445, + 0.008659828454256058, + 0.007811793591827154, + 0.006623103749006987, + 0.023832736536860466, + 0.0068060969933867455, + 0.023328939452767372, + 0.005897182039916515, + 0.004934969823807478, + 0.010564238764345646, + 0.0042802090756595135, + 0.005493987817317247, + 0.009174357168376446, + 0.003985383547842503, + 0.008420281112194061, + 0.019333234056830406, + 0.008722260594367981, + 0.01761116087436676, + 0.0071432506665587425, + 0.007425351068377495, + 0.013128913938999176, + 0.008749297820031643, + 0.022654414176940918, + 0.007039363030344248, + 0.00783123541623354, + 0.009653262794017792, + 0.007674077525734901, + 0.009659776464104652, + 0.00789166521281004, + 0.009913716465234756, + 0.017590101808309555, + 0.02825978957116604, + 0.013714941218495369, + 0.04360339418053627, + 0.014695830643177032, + 0.01801619678735733, + 0.027235539630055428, + 0.02750398963689804, + 0.04790487140417099, + 0.01302510779350996, + 0.009579343721270561, + 0.018525850027799606, + 0.007445964030921459, + 0.010870031081140041, + 0.012064345180988312, + 0.011202758178114891, + 0.010836072266101837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.004634710028767586 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a female warrior as she runs from left to right, advancing across the vast battlefield with a powerful yet somewhat weighted demeanor. Her long, dark brown hair, braided into intricate braids, whips violently as she runs. Her face is stern, stained with dust and sweat. Her eyes gaze sharply ahead, her lips pursed, revealing an unyielding resolve. She clads in a suit of ancient, battle-worn armor, a blend of bronze and leather. The breastplate is faintly engraved with totems, and the shoulder pads and wrist guards are scarred and dented. Beneath the armor lies a crimson, coarse cloth kilt, its hems billowing with her movements. Her arms swing vigorously for balance, each step sinking deeply into the muddy, uneven ground beneath her feet, sending small splashes of mud flying. In the background, a dark, smoke-filled sky rises. Sporadic fires burn in the distance, and broken flags and scattered weapons litter the desolate battlefield. The entire image exhibits a distinct sense of retro film grain, with tones tending towards low-saturation sepia and gray, colors that seem faded through the years, and slight vignetting and irregular light leaks at the edges, which together create a strong sense of epic and tragic atmosphere.", + "seed": 4148, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_7/index_8789_seed_4148_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01414803189691156, + "vlm_logprob": -0.01414803189691156, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.01414803189691156, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025802189484238625 + ], + "opposite_logprob_tokens": [ + -3.7758021354675293 + ], + "logprob_reward": -0.025802189484238625, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025802189484238625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06007048487663269, + 0.04852687940001488, + 0.0319257415831089, + 0.041771192103624344, + 0.11062097549438477, + 0.18698741495609283, + 0.22116272151470184, + 0.1094946637749672, + 0.09005587548017502, + 0.09938400238752365 + ], + "credit_map": [ + 0.007864128798246384, + 0.011766555719077587, + 0.006173732690513134, + 0.007369162980467081, + 0.007251066155731678, + 0.005690045654773712, + 0.008052293211221695, + 0.005903501529246569, + 0.00671790586784482, + 0.0046052937395870686, + 0.009477123618125916, + 0.005369157064706087, + 0.004124132916331291, + 0.003729404415935278, + 0.005407435353845358, + 0.009096425026655197, + 0.004899769090116024, + 0.006295318249613047, + 0.00394747219979763, + 0.003366560908034444, + 0.003949334844946861, + 0.00224847043864429, + 0.00404554046690464, + 0.0031732781790196896, + 0.004590045195072889, + 0.006212619598954916, + 0.003550104796886444, + 0.00824180617928505, + 0.004234157502651215, + 0.00473357317969203, + 0.0046229176223278046, + 0.005585967563092709, + 0.013682937249541283, + 0.016241025179624557, + 0.013007969595491886, + 0.009571515023708344, + 0.018757788464426994, + 0.013478637672960758, + 0.01561287883669138, + 0.01026822254061699, + 0.023802336305379868, + 0.027478504925966263, + 0.019829852506518364, + 0.023104695603251457, + 0.024061379954218864, + 0.024484219029545784, + 0.028596265241503716, + 0.015630176290869713, + 0.031106991693377495, + 0.0222146138548851, + 0.023499224334955215, + 0.035481806844472885, + 0.032151926308870316, + 0.023597482591867447, + 0.02492271177470684, + 0.028187979012727737, + 0.025883683934807777, + 0.028180673718452454, + 0.014539843425154686, + 0.008193857967853546, + 0.0078058550134301186, + 0.008740552701056004, + 0.008192604407668114, + 0.007957590743899345, + 0.012953085824847221, + 0.026400702074170113, + 0.0060968585312366486, + 0.00972467940300703, + 0.007850286550819874, + 0.015115334652364254, + 0.006845421157777309, + 0.0050695110112428665, + 0.010565252043306828, + 0.01704620197415352, + 0.01255673635751009, + 0.012910541146993637, + 0.010968666523694992, + 0.014116762205958366, + 0.010735657066106796, + 0.010484191589057446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.025802189484238625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420646533370018 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.01420646533370018, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420646533370018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12202070653438568, + 0.06024939939379692, + 0.07706636190414429, + 0.18368251621723175, + 0.14087547361850739, + 0.07324349880218506, + 0.042843762785196304, + 0.10817905515432358, + 0.04593455418944359, + 0.14590470492839813 + ], + "credit_map": [ + 0.01233390998095274, + 0.017848068848252296, + 0.01926918886601925, + 0.013206269592046738, + 0.015904396772384644, + 0.01500433124601841, + 0.0185108482837677, + 0.009943687357008457, + 0.00969045702368021, + 0.0083528533577919, + 0.007017733529210091, + 0.006487937178462744, + 0.007086239755153656, + 0.005836062598973513, + 0.007484964095056057, + 0.0082931537181139, + 0.008924092166125774, + 0.008689193986356258, + 0.011204681359231472, + 0.015948954969644547, + 0.008152198977768421, + 0.005095068830996752, + 0.010359224863350391, + 0.00869294535368681, + 0.016149330884218216, + 0.020452674478292465, + 0.016191743314266205, + 0.04906473308801651, + 0.014191880822181702, + 0.014502563513815403, + 0.03342393413186073, + 0.01970566064119339, + 0.015081365592777729, + 0.018201205879449844, + 0.014219668693840504, + 0.011347749270498753, + 0.032120637595653534, + 0.015715518966317177, + 0.02408985048532486, + 0.010099471546709538, + 0.012437322176992893, + 0.011015123687684536, + 0.006870192009955645, + 0.00945489015430212, + 0.00987490639090538, + 0.007860186509788036, + 0.009693725034594536, + 0.006037149578332901, + 0.00605808524414897, + 0.004084268119186163, + 0.004215265158563852, + 0.005825175903737545, + 0.008269181475043297, + 0.004617155063897371, + 0.004843203816562891, + 0.004931429401040077, + 0.032645564526319504, + 0.027087725698947906, + 0.012066787108778954, + 0.006222885567694902, + 0.006298007909208536, + 0.00782912876456976, + 0.010772046633064747, + 0.005256919655948877, + 0.005370007362216711, + 0.007138486951589584, + 0.0052172900177538395, + 0.005319160409271717, + 0.007515831850469112, + 0.006731561850756407, + 0.00501056807115674, + 0.003631647676229477, + 0.016760749742388725, + 0.028894029557704926, + 0.019454075023531914, + 0.01959490031003952, + 0.017997603863477707, + 0.016533171758055687, + 0.015074723400175571, + 0.01159544475376606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.01420646533370018 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328469604253769 + ], + "opposite_logprob_tokens": [ + -3.773284673690796 + ], + "logprob_reward": -0.02328469604253769, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328469604253769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1260405033826828, + 0.056474316865205765, + 0.04526428133249283, + 0.04670710489153862, + 0.07504013925790787, + 0.13676875829696655, + 0.1475055068731308, + 0.06747438758611679, + 0.07328147441148758, + 0.22544348239898682 + ], + "credit_map": [ + 0.012265976518392563, + 0.01804281584918499, + 0.014182479120790958, + 0.019389821216464043, + 0.01679621823132038, + 0.015739547088742256, + 0.01872495748102665, + 0.010898691602051258, + 0.009889601729810238, + 0.006744178477674723, + 0.008880987763404846, + 0.006723044440150261, + 0.005683328956365585, + 0.007036833092570305, + 0.005266994703561068, + 0.006249351426959038, + 0.009062236174941063, + 0.005414128303527832, + 0.005630196072161198, + 0.004996362142264843, + 0.00406367052346468, + 0.003564250422641635, + 0.006138984113931656, + 0.006394457072019577, + 0.004895148798823357, + 0.006048284005373716, + 0.005806969478726387, + 0.008088556118309498, + 0.0049499389715492725, + 0.003994982223957777, + 0.007045668549835682, + 0.0058775609359145164, + 0.008610047399997711, + 0.013750435784459114, + 0.007937467657029629, + 0.007472170516848564, + 0.013153494335711002, + 0.008559439331293106, + 0.008570871315896511, + 0.006986220832914114, + 0.018946275115013123, + 0.020221063867211342, + 0.011645953170955181, + 0.01693284511566162, + 0.015805615112185478, + 0.02718258649110794, + 0.01584494300186634, + 0.010189500637352467, + 0.0176274161785841, + 0.014607078395783901, + 0.011197821237146854, + 0.017652040347456932, + 0.029823418706655502, + 0.022292498499155045, + 0.015375299379229546, + 0.018929924815893173, + 0.010795933194458485, + 0.011979456059634686, + 0.011234061792492867, + 0.0068871621042490005, + 0.00795671809464693, + 0.00833811704069376, + 0.006525472272187471, + 0.003757472848519683, + 0.009907768107950687, + 0.011463018134236336, + 0.006653622258454561, + 0.0062239221297204494, + 0.012425294145941734, + 0.008214973844587803, + 0.009376687929034233, + 0.009016193449497223, + 0.028939494863152504, + 0.03462281450629234, + 0.024852300062775612, + 0.030172066763043404, + 0.02588368020951748, + 0.04029623419046402, + 0.02374085783958435, + 0.016936054453253746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.02328469604253769 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007151191821321845 + ], + "opposite_logprob_tokens": [ + -7.250715255737305 + ], + "logprob_reward": -0.0007151191821321845, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007151191821321845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28161829710006714, + 0.11013942211866379, + 0.04560641571879387, + 0.030060360208153725, + 0.03986259922385216, + 0.08006838709115982, + 0.09373066574335098, + 0.1749677211046219, + 0.0855761468410492, + 0.05837000533938408 + ], + "credit_map": [ + 0.02541792392730713, + 0.029047127813100815, + 0.055097371339797974, + 0.020420240238308907, + 0.0370003879070282, + 0.0333322249352932, + 0.048030704259872437, + 0.03327231481671333, + 0.010165094397962093, + 0.007411953527480364, + 0.01686403900384903, + 0.008852347731590271, + 0.012201391160488129, + 0.011448618955910206, + 0.015269781462848186, + 0.027926182374358177, + 0.008745115250349045, + 0.004194118082523346, + 0.00668628653511405, + 0.004605575930327177, + 0.004020852968096733, + 0.00311620463617146, + 0.007239943835884333, + 0.006998313590884209, + 0.0036795695777982473, + 0.00398587342351675, + 0.0030883760191500187, + 0.006936803460121155, + 0.0028848221991211176, + 0.0021891158539801836, + 0.0032054788898676634, + 0.004090318456292152, + 0.004954037722200155, + 0.005705251824110746, + 0.005538511089980602, + 0.003748090472072363, + 0.004389594774693251, + 0.0039506880566477776, + 0.007365978788584471, + 0.0042104413732886314, + 0.011628163047134876, + 0.013429502956569195, + 0.008875596337020397, + 0.010798980481922626, + 0.009446472860872746, + 0.01013097446411848, + 0.009849105961620808, + 0.005909583996981382, + 0.011785264126956463, + 0.009221045300364494, + 0.008266243152320385, + 0.013790642842650414, + 0.01601947844028473, + 0.011983733624219894, + 0.010747133754193783, + 0.01191711612045765, + 0.05844724550843239, + 0.05955326929688454, + 0.014006237499415874, + 0.00822908990085125, + 0.009525288827717304, + 0.011891430243849754, + 0.008096463046967983, + 0.005218674894422293, + 0.013773875311017036, + 0.027750102803111076, + 0.005193143151700497, + 0.0069408356212079525, + 0.007138903718441725, + 0.015463585034012794, + 0.005156356375664473, + 0.004159332253038883, + 0.005826893262565136, + 0.008506910875439644, + 0.008751444518566132, + 0.013791939243674278, + 0.004796996712684631, + 0.007070667110383511, + 0.00503645371645689, + 0.0045886896550655365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0007151191821321845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731689441949129 + ], + "opposite_logprob_tokens": [ + -5.0067315101623535 + ], + "logprob_reward": -0.006731689441949129, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731689441949129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15553635358810425, + 0.0701686292886734, + 0.0591556541621685, + 0.05054440349340439, + 0.1385214626789093, + 0.13912580907344818, + 0.07299389690160751, + 0.14417502284049988, + 0.0742616206407547, + 0.09551718831062317 + ], + "credit_map": [ + 0.015187753364443779, + 0.022218771278858185, + 0.012678836472332478, + 0.011964185163378716, + 0.020748155191540718, + 0.024268321692943573, + 0.03269115835428238, + 0.015779169276356697, + 0.009634626097977161, + 0.006583598908036947, + 0.009007158689200878, + 0.010718864388763905, + 0.008656343445181847, + 0.008295518346130848, + 0.009406779892742634, + 0.007865739054977894, + 0.009349268861114979, + 0.008375418372452259, + 0.007743224035948515, + 0.008830569684505463, + 0.005607691593468189, + 0.004044919740408659, + 0.008256576023995876, + 0.00694798119366169, + 0.0057831029407680035, + 0.00617171498015523, + 0.006070609204471111, + 0.01035030372440815, + 0.004666812252253294, + 0.004450446460396051, + 0.006232275627553463, + 0.006819138769060373, + 0.01950063928961754, + 0.02185225300490856, + 0.021886149421334267, + 0.01291215606033802, + 0.017254237085580826, + 0.014669202268123627, + 0.014363213442265987, + 0.016083603724837303, + 0.02004203014075756, + 0.023465722799301147, + 0.02223541960120201, + 0.019250039011240005, + 0.013903823681175709, + 0.016469014808535576, + 0.015495388768613338, + 0.008264374919235706, + 0.010839961469173431, + 0.007265541702508926, + 0.0048471475020051, + 0.007814101874828339, + 0.01693364605307579, + 0.010358593426644802, + 0.008511167019605637, + 0.006423736456781626, + 0.04517674446105957, + 0.04475059360265732, + 0.01119320746511221, + 0.00854422152042389, + 0.00863749161362648, + 0.01173680555075407, + 0.007508524693548679, + 0.006627434864640236, + 0.00974234938621521, + 0.018564099445939064, + 0.007058851420879364, + 0.008558903820812702, + 0.006232406944036484, + 0.0113367335870862, + 0.0066446871496737, + 0.006123591214418411, + 0.014450088143348694, + 0.02510594204068184, + 0.013951506465673447, + 0.012887267395853996, + 0.006347436457872391, + 0.008319462649524212, + 0.007392738945782185, + 0.007062750868499279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.006731689441949129 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a female warrior as she runs from left to right, advancing across the vast battlefield with a powerful yet somewhat weighted demeanor. Her long, dark brown hair, braided into intricate braids, whips violently as she runs. Her face is stern, stained with dust and sweat. Her eyes gaze sharply ahead, her lips pursed, revealing an unyielding resolve. She clads in a suit of ancient, battle-worn armor, a blend of bronze and leather. The breastplate is faintly engraved with totems, and the shoulder pads and wrist guards are scarred and dented. Beneath the armor lies a crimson, coarse cloth kilt, its hems billowing with her movements. Her arms swing vigorously for balance, each step sinking deeply into the muddy, uneven ground beneath her feet, sending small splashes of mud flying. In the background, a dark, smoke-filled sky rises. Sporadic fires burn in the distance, and broken flags and scattered weapons litter the desolate battlefield. The entire image exhibits a distinct sense of retro film grain, with tones tending towards low-saturation sepia and gray, colors that seem faded through the years, and slight vignetting and irregular light leaks at the edges, which together create a strong sense of epic and tragic atmosphere.", + "seed": 4149, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_7/index_8789_seed_4149_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0169319937704131, + "vlm_logprob": -0.0169319937704131, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.0169319937704131, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051856812089681625 + ], + "opposite_logprob_tokens": [ + -3.051856756210327 + ], + "logprob_reward": -0.051856812089681625, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.051856812089681625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07288329303264618, + 0.044720906764268875, + 0.04405463486909866, + 0.05556317791342735, + 0.20617078244686127, + 0.06328201293945312, + 0.08663337677717209, + 0.11175757646560669, + 0.1737072914838791, + 0.14122694730758667 + ], + "credit_map": [ + 0.010243757627904415, + 0.006535300053656101, + 0.012570388615131378, + 0.01335945539176464, + 0.007976125925779343, + 0.004528879653662443, + 0.011175350286066532, + 0.00649403128772974, + 0.005389441270381212, + 0.00531062250956893, + 0.007072456646710634, + 0.004409559536725283, + 0.006100797094404697, + 0.004136773757636547, + 0.007157043553888798, + 0.005144212394952774, + 0.0051709916442632675, + 0.006927290465682745, + 0.006368674803525209, + 0.0051838853396475315, + 0.0038148059975355864, + 0.007422150112688541, + 0.004723802674561739, + 0.0044430335983633995, + 0.007134153041988611, + 0.007421750109642744, + 0.008786479011178017, + 0.006313197314739227, + 0.007847599685192108, + 0.004892978351563215, + 0.007192334625869989, + 0.005974682047963142, + 0.056711599230766296, + 0.03546532988548279, + 0.011293478310108185, + 0.012510380707681179, + 0.017622241750359535, + 0.050306543707847595, + 0.010404913686215878, + 0.011856287717819214, + 0.007552777882665396, + 0.007331319153308868, + 0.007148542907088995, + 0.0065449741668999195, + 0.0104606281965971, + 0.007918095216155052, + 0.009224043227732182, + 0.007101637776941061, + 0.008562909439206123, + 0.013260036706924438, + 0.012927914038300514, + 0.012128490023314953, + 0.00816376507282257, + 0.010553603060543537, + 0.010709063149988651, + 0.010327599011361599, + 0.011792332865297794, + 0.010048643685877323, + 0.01389771793037653, + 0.010052735917270184, + 0.010651746764779091, + 0.020322365686297417, + 0.013633029535412788, + 0.02135901153087616, + 0.013785683549940586, + 0.02016567438840866, + 0.03678951784968376, + 0.017428753897547722, + 0.017770955339074135, + 0.02462862804532051, + 0.018447577953338623, + 0.02469051443040371, + 0.013438573107123375, + 0.009880829602479935, + 0.02020077593624592, + 0.012721938081085682, + 0.012924625538289547, + 0.02695428766310215, + 0.01153573114424944, + 0.033570174127817154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.051856812089681625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551716528832912 + ], + "opposite_logprob_tokens": [ + -4.387551784515381 + ], + "logprob_reward": -0.012551716528832912, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551716528832912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14475451409816742, + 0.10630513727664948, + 0.1535193771123886, + 0.11069299280643463, + 0.09612472355365753, + 0.10531241446733475, + 0.06297434121370316, + 0.08206463605165482, + 0.049801576882600784, + 0.08845028281211853 + ], + "credit_map": [ + 0.01803426630795002, + 0.013153845444321632, + 0.02124449610710144, + 0.017970068380236626, + 0.021904602646827698, + 0.01389247551560402, + 0.025776442140340805, + 0.01277831755578518, + 0.011191987432539463, + 0.01629354991018772, + 0.011851978488266468, + 0.009973736479878426, + 0.018877902999520302, + 0.007138834334909916, + 0.023111794143915176, + 0.007865347899496555, + 0.015294860117137432, + 0.017012933269143105, + 0.02208876982331276, + 0.014888497069478035, + 0.014200598001480103, + 0.027504578232765198, + 0.023067578673362732, + 0.019461560994386673, + 0.01685495302081108, + 0.010635068640112877, + 0.014010950922966003, + 0.010479089803993702, + 0.011034649796783924, + 0.012685433961451054, + 0.013420638628304005, + 0.021572209894657135, + 0.017821889370679855, + 0.016694458201527596, + 0.009096934460103512, + 0.013002690859138966, + 0.008849387057125568, + 0.009824821725487709, + 0.010397002100944519, + 0.010437545366585255, + 0.011799413710832596, + 0.010774384252727032, + 0.014704099856317043, + 0.010987715795636177, + 0.019221512600779533, + 0.013845295645296574, + 0.012849001213908195, + 0.011130984872579575, + 0.007411810103803873, + 0.007114718202501535, + 0.01248374953866005, + 0.004381385166198015, + 0.0034565033856779337, + 0.010890536941587925, + 0.005574287846684456, + 0.01166134886443615, + 0.008062916807830334, + 0.007683005649596453, + 0.007962990552186966, + 0.010815372690558434, + 0.007180286571383476, + 0.01248624362051487, + 0.009472095407545567, + 0.018401727080345154, + 0.0057344273664057255, + 0.006064273416996002, + 0.009273428469896317, + 0.005017937161028385, + 0.005401713773608208, + 0.006780729629099369, + 0.006295390427112579, + 0.005233677104115486, + 0.00929173268377781, + 0.008644742891192436, + 0.011995782144367695, + 0.008570650592446327, + 0.008342832326889038, + 0.014303598552942276, + 0.009789588861167431, + 0.01751134917140007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.012551716528832912 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419494766741991 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.01419494766741991, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419494766741991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10676110535860062, + 0.07592621445655823, + 0.10804186016321182, + 0.06527942419052124, + 0.16831104457378387, + 0.033367086201906204, + 0.05746044963598251, + 0.17386864125728607, + 0.17208652198314667, + 0.038897641003131866 + ], + "credit_map": [ + 0.016089577227830887, + 0.011479247361421585, + 0.017047349363565445, + 0.012535094283521175, + 0.011652287095785141, + 0.008038733154535294, + 0.02044452726840973, + 0.009474295191466808, + 0.010145281441509724, + 0.015489119105041027, + 0.009403139352798462, + 0.007790491450577974, + 0.009924504905939102, + 0.004493420012295246, + 0.011085961945354939, + 0.007594292983412743, + 0.013825079426169395, + 0.015429792925715446, + 0.02107236720621586, + 0.011055086739361286, + 0.009153197519481182, + 0.017033854499459267, + 0.008759853430092335, + 0.011712622828781605, + 0.011127760633826256, + 0.008091581985354424, + 0.00647558132186532, + 0.005465473514050245, + 0.0059857419691979885, + 0.010503120720386505, + 0.005801000166684389, + 0.011829162016510963, + 0.04088252782821655, + 0.030594147741794586, + 0.011080469004809856, + 0.01310682762414217, + 0.014498017728328705, + 0.03802832216024399, + 0.009089682251214981, + 0.011031057685613632, + 0.0041747139766812325, + 0.003903978504240513, + 0.0040631587617099285, + 0.003910684492439032, + 0.00587856862694025, + 0.004849154502153397, + 0.003579657059162855, + 0.003007171442732215, + 0.006304455455392599, + 0.0040871151722967625, + 0.009825102053582668, + 0.006679416634142399, + 0.004942529369145632, + 0.00769568607211113, + 0.0057173436507582664, + 0.012208801694214344, + 0.01826235093176365, + 0.02335054986178875, + 0.017394673079252243, + 0.018501877784729004, + 0.017826585099101067, + 0.023935945704579353, + 0.02222376875579357, + 0.03237288445234299, + 0.013116603717207909, + 0.01825786381959915, + 0.03741476684808731, + 0.0183193851262331, + 0.02214730530977249, + 0.02321276254951954, + 0.015104260295629501, + 0.024513578042387962, + 0.004317372106015682, + 0.004642433486878872, + 0.005605933256447315, + 0.0038896631449460983, + 0.004588205832988024, + 0.005327969789505005, + 0.0037980724591761827, + 0.006727990228682756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.01419494766741991 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008086272282525897 + ], + "opposite_logprob_tokens": [ + -7.1258087158203125 + ], + "logprob_reward": -0.0008086272282525897, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008086272282525897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1123417392373085, + 0.08269036561250687, + 0.1881118267774582, + 0.044308312237262726, + 0.24438656866550446, + 0.058502890169620514, + 0.10125276446342468, + 0.10185970366001129, + 0.03145131468772888, + 0.03509454429149628 + ], + "credit_map": [ + 0.01934082619845867, + 0.010796942748129368, + 0.0208441149443388, + 0.02105678617954254, + 0.009147215634584427, + 0.007417714688926935, + 0.01026713103055954, + 0.013471005484461784, + 0.01371344830840826, + 0.02014879137277603, + 0.012792346067726612, + 0.00676721939817071, + 0.0063799899071455, + 0.008833276107907295, + 0.006356362719088793, + 0.007698938716202974, + 0.029729096218943596, + 0.035538021475076675, + 0.02290305308997631, + 0.015152683481574059, + 0.007993272505700588, + 0.05520173907279968, + 0.010077477432787418, + 0.011516481637954712, + 0.004874395206570625, + 0.005297424737364054, + 0.007198317442089319, + 0.005732192657887936, + 0.006044681183993816, + 0.004892979748547077, + 0.005829768721014261, + 0.004438551142811775, + 0.058917101472616196, + 0.037312477827072144, + 0.01098165288567543, + 0.015714431181550026, + 0.022516468539834023, + 0.06506004184484482, + 0.016380511224269867, + 0.017503881826996803, + 0.006896498147398233, + 0.006382190156728029, + 0.008809346705675125, + 0.006126254331320524, + 0.00822829082608223, + 0.007973477244377136, + 0.005095123779028654, + 0.008991708047688007, + 0.012493688613176346, + 0.008319410495460033, + 0.017724979668855667, + 0.006517881527543068, + 0.005825422238558531, + 0.01506310049444437, + 0.008389969356358051, + 0.026918310672044754, + 0.010584455914795399, + 0.008099994622170925, + 0.010816509835422039, + 0.012516837567090988, + 0.01064437534660101, + 0.017517387866973877, + 0.010520798154175282, + 0.02115933783352375, + 0.0033118438441306353, + 0.0035279341973364353, + 0.00559103162959218, + 0.0036982207093387842, + 0.0036797255743294954, + 0.0043984330259263515, + 0.0033362179528921843, + 0.0039079091511666775, + 0.0038983372505754232, + 0.003734108991920948, + 0.00525594362989068, + 0.003510821843519807, + 0.0036677494645118713, + 0.006392833311110735, + 0.002736565424129367, + 0.005898181814700365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0008086272282525897 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a female warrior running from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the warrior wearing ancient battle-worn armor made of bronze and leather with a crimson kilt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the warrior's long dark brown hair braided and whipping violently as she runs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield with a dark smoke-filled sky and distant fires?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle tracking shot with a retro film grain aesthetic and low-saturation sepia tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247865337878466 + ], + "opposite_logprob_tokens": [ + -5.255248069763184 + ], + "logprob_reward": -0.005247865337878466, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247865337878466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1680523008108139, + 0.057374924421310425, + 0.11706158518791199, + 0.05026198923587799, + 0.09740875661373138, + 0.04086589068174362, + 0.06480765342712402, + 0.04845821484923363, + 0.22643017768859863, + 0.12927855551242828 + ], + "credit_map": [ + 0.04065578430891037, + 0.02093384601175785, + 0.02392200008034706, + 0.02144097536802292, + 0.01944274827837944, + 0.009045529179275036, + 0.01789700612425804, + 0.014714411459863186, + 0.008065625093877316, + 0.012321333400905132, + 0.009072422049939632, + 0.004767365287989378, + 0.006941408850252628, + 0.0064797052182257175, + 0.004634247627109289, + 0.005092816427350044, + 0.01849023997783661, + 0.019141066819429398, + 0.01366412453353405, + 0.008105846121907234, + 0.004789488390088081, + 0.04037712141871452, + 0.005730204284191132, + 0.0067634969018399715, + 0.006840065587311983, + 0.004953443072736263, + 0.010177980177104473, + 0.00700213061645627, + 0.006199117749929428, + 0.004714028909802437, + 0.006148187443614006, + 0.004227036144584417, + 0.02247077226638794, + 0.016131730750203133, + 0.00899568386375904, + 0.007238734979182482, + 0.008143248967826366, + 0.016335101798176765, + 0.010251741856336594, + 0.007841737940907478, + 0.006133324466645718, + 0.004150609485805035, + 0.00480609992519021, + 0.005230018403381109, + 0.0036385173443704844, + 0.005068858154118061, + 0.0040131802670657635, + 0.007825284264981747, + 0.0076453229412436485, + 0.008631504140794277, + 0.008124446496367455, + 0.00541177112609148, + 0.010868650861084461, + 0.009888896718621254, + 0.006653523072600365, + 0.007583537604659796, + 0.0048620328307151794, + 0.004792102612555027, + 0.005567906890064478, + 0.00588328018784523, + 0.004595121834427118, + 0.0062153828330338, + 0.005091317463666201, + 0.011451071128249168, + 0.019038978964090347, + 0.02415529452264309, + 0.04890479892492294, + 0.021740727126598358, + 0.026224594563245773, + 0.03307940438389778, + 0.019272729754447937, + 0.034013647586107254, + 0.014111893251538277, + 0.010593339800834656, + 0.01659271866083145, + 0.011032589711248875, + 0.015209231525659561, + 0.022701170295476913, + 0.012388093397021294, + 0.02664951980113983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.005247865337878466 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_8/index_tensor([3182, 3182])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_8/index_tensor([3182, 3182])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6e70db0e39dcc671866566216deb5334991606b5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_8/index_tensor([3182, 3182])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "locked, completely still long shot gazes steadily at a nearly vertical, weathered, gray rock face that dominates the frame. To one side, a massive waterfall cascades down from the heights like a galaxy, crashing into the deep pool below, stirring up a cloud of dancing mist. This dense mist envelops the lower half of the scene, slowly curling and rising in front of the cliff. Cerebral light filters through thick clouds and pervasive vapor, transforming into a soft, evenly diffused glow, gently illuminating every detail of the rock face. On this natural, monolithic canvas, the majestic figures of the Four Heavenly Kings are carved in massive relief. Clad in intricate armor and wielding their iconic magical instruments, their faces stern and resolute, their eyes wide open, their expressions solemn and majestic. The surface of the statue is not smooth stone, but rather exhibits the rich, granular texture characteristic of rock painting. Its vibrant and quaint colors, richly applied in cinnabar red, azurite, malachite green, and ochre, appear exceptionally saturated and profound in the humid air. Countless tiny water droplets cling to the statue's mottled hues, shimmering and adding a touch of vividness to the otherwise tranquil scene. The entire painting is imbued with a sense of grandeur, mystery, and sacred power, creating an awe-inspiring atmosphere of tranquility.", + "seed": 6780, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_8/index_3182_seed_6780_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03718013232573867, + "vlm_logprob": -0.03718013232573867, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03718013232573867, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12823432683944702 + ], + "opposite_logprob_tokens": [ + -2.253234386444092 + ], + "logprob_reward": -0.12823432683944702, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12823432683944702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09667865931987762, + 0.007163597270846367, + 0.018944188952445984, + 0.012442589737474918, + 0.007516184821724892, + 0.009794044308364391, + 0.7013148069381714, + 0.011851709336042404, + 0.11967483907938004, + 0.014619321562349796 + ], + "credit_map": [ + 0.00804164633154869, + 0.004490458406507969, + 0.02646375261247158, + 0.0059904493391513824, + 0.0024334874469786882, + 0.014598536305129528, + 0.00756295770406723, + 0.027097374200820923, + 0.0006636264733970165, + 0.0006762493867427111, + 0.0018207032699137926, + 0.0007081280928105116, + 0.0006444854661822319, + 0.0007800915627740324, + 0.0008907412411645055, + 0.000979571370407939, + 0.0018040037248283625, + 0.000983166741207242, + 0.0027185543440282345, + 0.0015655136667191982, + 0.0009905012557283044, + 0.002823679009452462, + 0.0044187018647789955, + 0.003640067530795932, + 0.0009692980092950165, + 0.001709446427412331, + 0.0012409965274855494, + 0.0018570426618680358, + 0.001414748141542077, + 0.0017121384153142571, + 0.0013215311337262392, + 0.0022173889447003603, + 0.0009553605341352522, + 0.0005486828740686178, + 0.0010797783033922315, + 0.0006660139188170433, + 0.00047171907499432564, + 0.0011847063433378935, + 0.001126919756643474, + 0.0014830041909590364, + 0.0009408621117472649, + 0.0009502103785052896, + 0.0013089810963720083, + 0.000867768656462431, + 0.0009735786006785929, + 0.001424696296453476, + 0.0016823060577735305, + 0.0016456420999020338, + 0.10452382266521454, + 0.022363802418112755, + 0.26487356424331665, + 0.04001092165708542, + 0.017245449125766754, + 0.16794945299625397, + 0.04573240503668785, + 0.0386153906583786, + 0.0015246925177052617, + 0.0013892619172111154, + 0.00106040108948946, + 0.001354027888737619, + 0.001077490160241723, + 0.0015479665016755462, + 0.0018127409275621176, + 0.002085127867758274, + 0.01409673411399126, + 0.007580946199595928, + 0.027644604444503784, + 0.008051231503486633, + 0.004932600539177656, + 0.015024257823824883, + 0.00884117092937231, + 0.03350329399108887, + 0.0013764941832050681, + 0.001541912672109902, + 0.0023165408056229353, + 0.001808258006349206, + 0.0011743531795218587, + 0.0021523069590330124, + 0.0014712829142808914, + 0.002778172492980957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.12823432683944702 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004112122114747763 + ], + "opposite_logprob_tokens": [ + -5.504112243652344 + ], + "logprob_reward": -0.004112122114747763, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004112122114747763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3870106637477875, + 0.24064573645591736, + 0.03604071959853172, + 0.013576246798038483, + 0.015848636627197266, + 0.023294080048799515, + 0.14050650596618652, + 0.019094308838248253, + 0.09239785373210907, + 0.03158527985215187 + ], + "credit_map": [ + 0.03310960903763771, + 0.015332252718508244, + 0.11095378547906876, + 0.022711506113409996, + 0.00920957326889038, + 0.05863199010491371, + 0.02779366262257099, + 0.10926828533411026, + 0.0230619665235281, + 0.010258304886519909, + 0.07477757334709167, + 0.018801063299179077, + 0.008191373199224472, + 0.027194440364837646, + 0.024439457803964615, + 0.053921546787023544, + 0.0036722379736602306, + 0.002107289619743824, + 0.00717361643910408, + 0.0027083484455943108, + 0.0017474540509283543, + 0.0053142374381423, + 0.006338305771350861, + 0.006979229860007763, + 0.0010067227995023131, + 0.001583868288435042, + 0.0013986582634970546, + 0.0018319443333894014, + 0.001505689462646842, + 0.0022822304163128138, + 0.0014355479506775737, + 0.002531585516408086, + 0.0014877073699608445, + 0.001465491484850645, + 0.002530468860641122, + 0.0021856233943253756, + 0.0014779496705159545, + 0.0015936355339363217, + 0.0024433976504951715, + 0.00266436324454844, + 0.0017445591511204839, + 0.0025052062701433897, + 0.00219710823148489, + 0.0024222105275839567, + 0.001967686926946044, + 0.003794380696490407, + 0.0034825822804123163, + 0.005180345848202705, + 0.020520681515336037, + 0.004719620570540428, + 0.05236733704805374, + 0.008231074549257755, + 0.0035245397593826056, + 0.030575748533010483, + 0.009288198314607143, + 0.011279311962425709, + 0.0023966748267412186, + 0.001659476081840694, + 0.0021775870118290186, + 0.0022952992003411055, + 0.0013701216084882617, + 0.0025801830925047398, + 0.0026052577886730433, + 0.004009709227830172, + 0.010411029681563377, + 0.005637782625854015, + 0.021908968687057495, + 0.005931819789111614, + 0.003527014749124646, + 0.011591734364628792, + 0.006050416734069586, + 0.027339093387126923, + 0.0028120067436248064, + 0.0028838017024099827, + 0.00560308713465929, + 0.0040700240060687065, + 0.0022980370558798313, + 0.004431383218616247, + 0.002925759181380272, + 0.006561180576682091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.004112122114747763 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016082797199487686 + ], + "opposite_logprob_tokens": [ + -4.141082763671875 + ], + "logprob_reward": -0.016082797199487686, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016082797199487686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2922438085079193, + 0.1385035663843155, + 0.050046566873788834, + 0.016355648636817932, + 0.03391055017709732, + 0.044715024530887604, + 0.30054715275764465, + 0.02760927565395832, + 0.05068565160036087, + 0.04538271576166153 + ], + "credit_map": [ + 0.025679469108581543, + 0.01592119224369526, + 0.07878925651311874, + 0.019338414072990417, + 0.013890661299228668, + 0.043800659477710724, + 0.023357875645160675, + 0.07146632671356201, + 0.011423916555941105, + 0.006793863605707884, + 0.04193057492375374, + 0.009641733020544052, + 0.005905953701585531, + 0.016244148835539818, + 0.013912942260503769, + 0.03265044838190079, + 0.004339313600212336, + 0.0035532601177692413, + 0.010985305532813072, + 0.003776068100705743, + 0.00303515768609941, + 0.005751091055572033, + 0.005032374989241362, + 0.01357400231063366, + 0.001263075857423246, + 0.0018842366989701986, + 0.00163913588039577, + 0.0025475365109741688, + 0.0016370394732803106, + 0.0023554253857582808, + 0.0019073624862357974, + 0.0031218386720865965, + 0.0029267477802932262, + 0.004846365191042423, + 0.003829122753813863, + 0.004078176338225603, + 0.004880622494965792, + 0.003652798943221569, + 0.004753893241286278, + 0.004942824598401785, + 0.0027702110819518566, + 0.004686594475060701, + 0.003503402229398489, + 0.004270531702786684, + 0.004386910703033209, + 0.007586709689348936, + 0.00694823544472456, + 0.010562431998550892, + 0.04455889016389847, + 0.009711196646094322, + 0.11158788949251175, + 0.017180176451802254, + 0.007712674327194691, + 0.0752047523856163, + 0.0196270402520895, + 0.014964531175792217, + 0.003341778414323926, + 0.0035936227068305016, + 0.002811003476381302, + 0.0025985068641602993, + 0.002709587337449193, + 0.004344527143985033, + 0.0038339735474437475, + 0.0043762787245213985, + 0.005476877558976412, + 0.004361813422292471, + 0.008263565599918365, + 0.004561085253953934, + 0.003854049602523446, + 0.008165363222360611, + 0.006023808848112822, + 0.009979091584682465, + 0.004067166708409786, + 0.0038480174262076616, + 0.007204374298453331, + 0.006953401956707239, + 0.0035737110301852226, + 0.006398195866495371, + 0.004973116796463728, + 0.008364732377231121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.016082797199487686 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014199530705809593 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.014199530705809593, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014199530705809593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3026820719242096, + 0.05495356395840645, + 0.04058723524212837, + 0.013534723781049252, + 0.04106263071298599, + 0.052834585309028625, + 0.37908703088760376, + 0.023439960554242134, + 0.03589377552270889, + 0.055924441665410995 + ], + "credit_map": [ + 0.02516094222664833, + 0.012661186046898365, + 0.08216887712478638, + 0.020457344129681587, + 0.007964321412146091, + 0.04693780094385147, + 0.02170470543205738, + 0.08562690019607544, + 0.0041787344962358475, + 0.004387912806123495, + 0.013306663371622562, + 0.0047102211974561214, + 0.004193402826786041, + 0.006590237375348806, + 0.005404177121818066, + 0.012182212434709072, + 0.0037174595054239035, + 0.003579393494874239, + 0.008431002497673035, + 0.003616916248574853, + 0.0029667827766388655, + 0.005557062104344368, + 0.005119241774082184, + 0.007599376607686281, + 0.0013748863711953163, + 0.0012993161799386144, + 0.0031262985430657864, + 0.0016119012143462896, + 0.0012284446274861693, + 0.0015535018173977733, + 0.0013963737292215228, + 0.001944000949151814, + 0.003624152159318328, + 0.005883868783712387, + 0.004954802803695202, + 0.005885482765734196, + 0.005658184178173542, + 0.004027575254440308, + 0.005245563108474016, + 0.005783004220575094, + 0.004260007292032242, + 0.008962482213973999, + 0.005485340021550655, + 0.006549467798322439, + 0.006998962257057428, + 0.0074564144015312195, + 0.006016661413013935, + 0.00710524944588542, + 0.05638667568564415, + 0.012185356579720974, + 0.14272427558898926, + 0.021486276760697365, + 0.009552708826959133, + 0.09209457039833069, + 0.025095421820878983, + 0.01956173963844776, + 0.0021206310484558344, + 0.0037458790466189384, + 0.0025775812100619078, + 0.003094028914347291, + 0.00273308833129704, + 0.0033145600464195013, + 0.002481592120602727, + 0.0033725991379469633, + 0.00399984372779727, + 0.002117094350978732, + 0.007511346600949764, + 0.0027317549102008343, + 0.0020646899938583374, + 0.005152345634996891, + 0.0035232165828347206, + 0.008793486282229424, + 0.00471665570512414, + 0.004898172803223133, + 0.008798103779554367, + 0.008022308349609375, + 0.004195323679596186, + 0.007856342010200024, + 0.005963712465018034, + 0.0114738205447793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.014199530705809593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327188476920128 + ], + "opposite_logprob_tokens": [ + -3.7732717990875244 + ], + "logprob_reward": -0.02327188476920128, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327188476920128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17746397852897644, + 0.07410930097103119, + 0.05485326424241066, + 0.016678359359502792, + 0.029984256252646446, + 0.026236381381750107, + 0.5697832703590393, + 0.02629176713526249, + 0.010055548511445522, + 0.014543889090418816 + ], + "credit_map": [ + 0.015051368623971939, + 0.007037387229502201, + 0.049591079354286194, + 0.010915151797235012, + 0.004544971976429224, + 0.026685593649744987, + 0.013709821738302708, + 0.049928609281778336, + 0.006048760376870632, + 0.0032920523080974817, + 0.022554747760295868, + 0.004957700613886118, + 0.0019368709763512015, + 0.009481712244451046, + 0.007675818633288145, + 0.01816163770854473, + 0.0065369983203709126, + 0.0035347994416952133, + 0.016261715441942215, + 0.004181393887847662, + 0.0027856347151100636, + 0.010569879785180092, + 0.0066545335575938225, + 0.004328310489654541, + 0.0015243777306750417, + 0.002009707735851407, + 0.0020088409073650837, + 0.0022439982276409864, + 0.001801149221137166, + 0.00244754645973444, + 0.0016768258064985275, + 0.002965912688523531, + 0.003228425281122327, + 0.0038070438895374537, + 0.004473804030567408, + 0.003539249300956726, + 0.0033950062934309244, + 0.003373760264366865, + 0.004447757266461849, + 0.0037192092277109623, + 0.003016773611307144, + 0.002797007793560624, + 0.004512109328061342, + 0.00248822127468884, + 0.001967697637155652, + 0.0040158601477742195, + 0.003647448727861047, + 0.0037912619300186634, + 0.08463799953460693, + 0.018137555569410324, + 0.2140347957611084, + 0.03239167109131813, + 0.014430259354412556, + 0.13988226652145386, + 0.03732287138700485, + 0.028945855796337128, + 0.003343723714351654, + 0.002142905490472913, + 0.002164102392271161, + 0.0026574842631816864, + 0.001639893394894898, + 0.004456197377294302, + 0.004304575268179178, + 0.005582884885370731, + 0.0011838076170533895, + 0.0009983166819438338, + 0.0015104023041203618, + 0.0008177738054655492, + 0.0008464214042760432, + 0.0015868733171373606, + 0.0011826271656900644, + 0.0019293265650048852, + 0.00141990021802485, + 0.0009182373178191483, + 0.002607583999633789, + 0.001784822205081582, + 0.0009540233877487481, + 0.0020947593729943037, + 0.0014463349943980575, + 0.003318228293210268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.02327188476920128 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "locked, completely still long shot gazes steadily at a nearly vertical, weathered, gray rock face that dominates the frame. To one side, a massive waterfall cascades down from the heights like a galaxy, crashing into the deep pool below, stirring up a cloud of dancing mist. This dense mist envelops the lower half of the scene, slowly curling and rising in front of the cliff. Cerebral light filters through thick clouds and pervasive vapor, transforming into a soft, evenly diffused glow, gently illuminating every detail of the rock face. On this natural, monolithic canvas, the majestic figures of the Four Heavenly Kings are carved in massive relief. Clad in intricate armor and wielding their iconic magical instruments, their faces stern and resolute, their eyes wide open, their expressions solemn and majestic. The surface of the statue is not smooth stone, but rather exhibits the rich, granular texture characteristic of rock painting. Its vibrant and quaint colors, richly applied in cinnabar red, azurite, malachite green, and ochre, appear exceptionally saturated and profound in the humid air. Countless tiny water droplets cling to the statue's mottled hues, shimmering and adding a touch of vividness to the otherwise tranquil scene. The entire painting is imbued with a sense of grandeur, mystery, and sacred power, creating an awe-inspiring atmosphere of tranquility.", + "seed": 6781, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_8/index_3182_seed_6781_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07875315789133311, + "vlm_logprob": -0.07875315789133311, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.07875315789133311, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31873199343681335 + ], + "opposite_logprob_tokens": [ + -1.3187320232391357 + ], + "logprob_reward": -0.31873199343681335, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31873199343681335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32746222615242004, + 0.2078157663345337, + 0.12954683601856232, + 0.07801716029644012, + 0.04041887819766998, + 0.030384158715605736, + 0.03676709532737732, + 0.029002951458096504, + 0.0478009432554245, + 0.07278399914503098 + ], + "credit_map": [ + 0.03537323325872421, + 0.04207329824566841, + 0.02432296611368656, + 0.050426166504621506, + 0.05110950767993927, + 0.044038768857717514, + 0.042653221637010574, + 0.03746504709124565, + 0.01947399415075779, + 0.03770402446389198, + 0.02434138022363186, + 0.04560314491391182, + 0.026346517726778984, + 0.014735051430761814, + 0.01890069991350174, + 0.020710919052362442, + 0.010196397081017494, + 0.02412230707705021, + 0.015049302950501442, + 0.0290965773165226, + 0.011384950019419193, + 0.010262088850140572, + 0.014818505384027958, + 0.01461668498814106, + 0.009063638746738434, + 0.013424525037407875, + 0.007449756376445293, + 0.01325021032243967, + 0.00783789437264204, + 0.004395618569105864, + 0.011433691717684269, + 0.011161815375089645, + 0.005054726731032133, + 0.004765825811773539, + 0.0036833956837654114, + 0.00481218658387661, + 0.006217021029442549, + 0.0025923659559339285, + 0.008212972432374954, + 0.005080382339656353, + 0.005052392370998859, + 0.004509117919951677, + 0.002913544187322259, + 0.0048491982743144035, + 0.0030468555632978678, + 0.001998597290366888, + 0.003569304244592786, + 0.004445143509656191, + 0.005963290575891733, + 0.004163017962127924, + 0.003711928613483906, + 0.005988996010273695, + 0.004121439531445503, + 0.003436698578298092, + 0.005543902516365051, + 0.0038378185126930475, + 0.005082220304757357, + 0.005444979760795832, + 0.0026030868757516146, + 0.0045839822851121426, + 0.0025162131059914827, + 0.0023590433411300182, + 0.002674409421160817, + 0.003739013569429517, + 0.00483427569270134, + 0.0047309682704508305, + 0.005123619921505451, + 0.005615119822323322, + 0.0046333740465343, + 0.0064988997764885426, + 0.005148047581315041, + 0.011216633953154087, + 0.01093654427677393, + 0.013391571119427681, + 0.008927205577492714, + 0.009778213687241077, + 0.007417448330670595, + 0.005534114316105843, + 0.006600739434361458, + 0.01019815169274807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.31873199343681335 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979840662330389 + ], + "opposite_logprob_tokens": [ + -4.634798526763916 + ], + "logprob_reward": -0.00979840662330389, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979840662330389, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3802114725112915, + 0.1439199298620224, + 0.12100417166948318, + 0.07608287036418915, + 0.05535208061337471, + 0.03799208998680115, + 0.05665573850274086, + 0.07077770680189133, + 0.01926497370004654, + 0.03873898833990097 + ], + "credit_map": [ + 0.03616131469607353, + 0.044899240136146545, + 0.05000557750463486, + 0.04114129766821861, + 0.06388366222381592, + 0.043198395520448685, + 0.0564475879073143, + 0.04447438567876816, + 0.01511282380670309, + 0.02022274024784565, + 0.019094934687018394, + 0.019568653777241707, + 0.021469784900546074, + 0.010713078081607819, + 0.020256327465176582, + 0.01748158410191536, + 0.01036890409886837, + 0.024304687976837158, + 0.01240486465394497, + 0.024104956537485123, + 0.016681089997291565, + 0.006129052489995956, + 0.015448015183210373, + 0.011562597006559372, + 0.011247780174016953, + 0.013337468728423119, + 0.006665170658379793, + 0.011354098096489906, + 0.00961995217949152, + 0.004812459461390972, + 0.010206800885498524, + 0.00883913692086935, + 0.008286244235932827, + 0.008179852738976479, + 0.0057003675028681755, + 0.009035839699208736, + 0.005567340180277824, + 0.0029596418607980013, + 0.00622741412371397, + 0.009395377710461617, + 0.004606423433870077, + 0.005428282078355551, + 0.0038497059140354395, + 0.006521313916891813, + 0.0043515395373106, + 0.0024284464307129383, + 0.00522975716739893, + 0.005576621741056442, + 0.007740121800452471, + 0.011679694056510925, + 0.00531366840004921, + 0.010633079335093498, + 0.005607519764453173, + 0.004106555134057999, + 0.005349243525415659, + 0.006225856952369213, + 0.00924723595380783, + 0.014778682962059975, + 0.006285190582275391, + 0.01186817605048418, + 0.0064277974888682365, + 0.005824497900903225, + 0.00675175292417407, + 0.009594377130270004, + 0.0022469405084848404, + 0.0038010706193745136, + 0.0018027627374976873, + 0.003091969760134816, + 0.0020255432464182377, + 0.0018497154815122485, + 0.0022061357740312815, + 0.0022408345248550177, + 0.005358233582228422, + 0.0038327842485159636, + 0.004723791498690844, + 0.005184629466384649, + 0.0031784246675670147, + 0.003942971583455801, + 0.004922396503388882, + 0.007595757022500038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00979840662330389 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079511493444443 + ], + "opposite_logprob_tokens": [ + -4.141079425811768 + ], + "logprob_reward": -0.016079511493444443, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079511493444443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41754356026649475, + 0.13992604613304138, + 0.10105223208665848, + 0.09927739948034286, + 0.08163508772850037, + 0.03980591520667076, + 0.02119198627769947, + 0.035305313766002655, + 0.020257357507944107, + 0.044005073606967926 + ], + "credit_map": [ + 0.03287870064377785, + 0.03607422113418579, + 0.033763688057661057, + 0.048722270876169205, + 0.08199991285800934, + 0.038719628006219864, + 0.10854530334472656, + 0.03683982789516449, + 0.010966083966195583, + 0.014127865433692932, + 0.012182571925222874, + 0.015551230870187283, + 0.029630698263645172, + 0.009687134064733982, + 0.030515193939208984, + 0.017265256494283676, + 0.013969913125038147, + 0.014217562973499298, + 0.008974536322057247, + 0.015032785944640636, + 0.014168906956911087, + 0.006409293971955776, + 0.019626125693321228, + 0.00865311361849308, + 0.01496429555118084, + 0.01585097424685955, + 0.009506604634225368, + 0.013171140104532242, + 0.013607132248580456, + 0.005402646493166685, + 0.01969439908862114, + 0.007080212701112032, + 0.009539454244077206, + 0.009582968428730965, + 0.006464907433837652, + 0.009988285601139069, + 0.0150192491710186, + 0.004498592112213373, + 0.01811089739203453, + 0.008430731482803822, + 0.005047393962740898, + 0.004026682581752539, + 0.0036206580698490143, + 0.005403790157288313, + 0.007097838446497917, + 0.0033827079460024834, + 0.007119632326066494, + 0.00410721218213439, + 0.0027459184639155865, + 0.002207504352554679, + 0.0019680571276694536, + 0.00304161524400115, + 0.0034023283515125513, + 0.0024156116414815187, + 0.0031854528933763504, + 0.0022254984360188246, + 0.004005859140306711, + 0.006143513601273298, + 0.002565177157521248, + 0.004604284651577473, + 0.005242676939815283, + 0.0036362684331834316, + 0.004700647201389074, + 0.004406889900565147, + 0.002981899306178093, + 0.0031391149386763573, + 0.0018271433655172586, + 0.0031286764424294233, + 0.0023814078886061907, + 0.0018014769302681088, + 0.0024659386835992336, + 0.0025316993705928326, + 0.005290449596941471, + 0.003617034526541829, + 0.005399874411523342, + 0.004973702132701874, + 0.004767019767314196, + 0.004582876339554787, + 0.0055382996797561646, + 0.009835819713771343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.016079511493444443 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075330898165703 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075330898165703, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075330898165703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6384367942810059, + 0.09345627576112747, + 0.07978473603725433, + 0.04369521886110306, + 0.03146196901798248, + 0.021425684913992882, + 0.01439847331494093, + 0.022067269310355186, + 0.015067100524902344, + 0.04020646587014198 + ], + "credit_map": [ + 0.04857555776834488, + 0.04943488538265228, + 0.05225958302617073, + 0.07578688114881516, + 0.12600840628147125, + 0.05411393940448761, + 0.17160622775554657, + 0.06065133586525917, + 0.009744576178491116, + 0.013671194203197956, + 0.012279989197850227, + 0.019623756408691406, + 0.010390742681920528, + 0.007200249005109072, + 0.010018536821007729, + 0.010527231730520725, + 0.0076599204912781715, + 0.015069058164954185, + 0.007788600865751505, + 0.01447746530175209, + 0.008509300649166107, + 0.0074737342074513435, + 0.010253829881548882, + 0.00855281949043274, + 0.005544919054955244, + 0.008684555068612099, + 0.0042718215845525265, + 0.006422694306820631, + 0.004860037937760353, + 0.0034653190523386, + 0.005916142836213112, + 0.004529728554189205, + 0.00642426498234272, + 0.004568806849420071, + 0.002804558025673032, + 0.004404548555612564, + 0.003920725546777248, + 0.0016890281112864614, + 0.003858537646010518, + 0.0037914973217993975, + 0.0026064461562782526, + 0.0023621809668838978, + 0.002248534234240651, + 0.003242767183110118, + 0.004024699330329895, + 0.0023417582269757986, + 0.0027101703453809023, + 0.0018891285872086883, + 0.0019303904846310616, + 0.0020107077434659004, + 0.0014296170556917787, + 0.001817195094190538, + 0.002629299182444811, + 0.0012343924026936293, + 0.002057658974081278, + 0.0012892112135887146, + 0.0024918855633586645, + 0.003037218702957034, + 0.001955975079908967, + 0.0029911857564002275, + 0.00407377490773797, + 0.0023368147667497396, + 0.0026189261116087437, + 0.0025614886544644833, + 0.0014668015064671636, + 0.00204077479429543, + 0.0012105026980862021, + 0.0020475389901548624, + 0.0022263380233198404, + 0.0019643171690404415, + 0.001662496360950172, + 0.002448330633342266, + 0.003864261554554105, + 0.004272610880434513, + 0.003983661066740751, + 0.004677875433117151, + 0.004202947020530701, + 0.005190737079828978, + 0.004169062245637178, + 0.009845309890806675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011075330898165703 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038080547004938126 + ], + "opposite_logprob_tokens": [ + -3.2880804538726807 + ], + "logprob_reward": -0.038080547004938126, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038080547004938126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3386293649673462, + 0.1654192954301834, + 0.07652026414871216, + 0.05028633028268814, + 0.0872313603758812, + 0.062306925654411316, + 0.03339206799864769, + 0.09911090135574341, + 0.021389897912740707, + 0.06571359932422638 + ], + "credit_map": [ + 0.024359870702028275, + 0.027575699612498283, + 0.026177890598773956, + 0.04035358503460884, + 0.06621220707893372, + 0.04166942834854126, + 0.08052632957696915, + 0.03175438195466995, + 0.014129789546132088, + 0.02798817679286003, + 0.018574893474578857, + 0.026450322940945625, + 0.02122584357857704, + 0.0147262467071414, + 0.02088245376944542, + 0.02144155651330948, + 0.009887391701340675, + 0.011693831533193588, + 0.00766547117382288, + 0.01391565427184105, + 0.009062071330845356, + 0.005650775041431189, + 0.010383917018771172, + 0.008261154405772686, + 0.00889006070792675, + 0.011008200235664845, + 0.004578860942274332, + 0.007719286251813173, + 0.005824091844260693, + 0.0024962800089269876, + 0.005132216960191727, + 0.0046373349614441395, + 0.011875415220856667, + 0.013002697378396988, + 0.007606260944157839, + 0.01834164746105671, + 0.008501040749251842, + 0.004571032244712114, + 0.010327368043363094, + 0.013005892746150494, + 0.009615058079361916, + 0.009852349758148193, + 0.005586997140198946, + 0.010155940428376198, + 0.005657693836838007, + 0.005089615471661091, + 0.007332478649914265, + 0.009016793221235275, + 0.00463899178430438, + 0.004540733993053436, + 0.003110535442829132, + 0.005236728582531214, + 0.0038437056355178356, + 0.002036381745710969, + 0.005002346821129322, + 0.004982643760740757, + 0.015373611822724342, + 0.016878221184015274, + 0.008264459669589996, + 0.01936572976410389, + 0.009407557547092438, + 0.008848534896969795, + 0.008462750352919102, + 0.01251003984361887, + 0.003047705627977848, + 0.0039047368336468935, + 0.0019462486961856484, + 0.00328409974463284, + 0.002324015600606799, + 0.0014291915576905012, + 0.0025220008101314306, + 0.0029318982269614935, + 0.008407398127019405, + 0.00601379107683897, + 0.008410140872001648, + 0.008085767738521099, + 0.005315006710588932, + 0.00764064397662878, + 0.00673571415245533, + 0.015105140395462513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.038080547004938126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "locked, completely still long shot gazes steadily at a nearly vertical, weathered, gray rock face that dominates the frame. To one side, a massive waterfall cascades down from the heights like a galaxy, crashing into the deep pool below, stirring up a cloud of dancing mist. This dense mist envelops the lower half of the scene, slowly curling and rising in front of the cliff. Cerebral light filters through thick clouds and pervasive vapor, transforming into a soft, evenly diffused glow, gently illuminating every detail of the rock face. On this natural, monolithic canvas, the majestic figures of the Four Heavenly Kings are carved in massive relief. Clad in intricate armor and wielding their iconic magical instruments, their faces stern and resolute, their eyes wide open, their expressions solemn and majestic. The surface of the statue is not smooth stone, but rather exhibits the rich, granular texture characteristic of rock painting. Its vibrant and quaint colors, richly applied in cinnabar red, azurite, malachite green, and ochre, appear exceptionally saturated and profound in the humid air. Countless tiny water droplets cling to the statue's mottled hues, shimmering and adding a touch of vividness to the otherwise tranquil scene. The entire painting is imbued with a sense of grandeur, mystery, and sacred power, creating an awe-inspiring atmosphere of tranquility.", + "seed": 6782, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_8/index_3182_seed_6782_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.026242625992745162, + "vlm_logprob": -0.026242625992745162, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.026242625992745162, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08639074862003326 + ], + "opposite_logprob_tokens": [ + -2.586390733718872 + ], + "logprob_reward": -0.08639074862003326, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08639074862003326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2521288990974426, + 0.12254488468170166, + 0.07286690175533295, + 0.05748508870601654, + 0.08021053671836853, + 0.16303583979606628, + 0.08228400349617004, + 0.09494096785783768, + 0.040458936244249344, + 0.03404396399855614 + ], + "credit_map": [ + 0.025448448956012726, + 0.03387654572725296, + 0.028192659839987755, + 0.0354570709168911, + 0.0500270277261734, + 0.02427958883345127, + 0.031240636482834816, + 0.023606907576322556, + 0.030684921890497208, + 0.00946724135428667, + 0.033534444868564606, + 0.011694395914673805, + 0.01325533352792263, + 0.007758048363029957, + 0.009363901801407337, + 0.006786591839045286, + 0.014225891791284084, + 0.006855999119579792, + 0.014975721947848797, + 0.006774011999368668, + 0.008456986397504807, + 0.006709452252835035, + 0.006885732989758253, + 0.007983105257153511, + 0.006818396970629692, + 0.00978588592261076, + 0.0055700563825666904, + 0.011752632446587086, + 0.005790778435766697, + 0.004313623998314142, + 0.007928954437375069, + 0.005524760112166405, + 0.007386140991002321, + 0.00955475214868784, + 0.008074170909821987, + 0.010959162376821041, + 0.010750159621238708, + 0.009157119318842888, + 0.013929198496043682, + 0.010399832390248775, + 0.013675859197974205, + 0.027567241340875626, + 0.019047876819968224, + 0.03062582202255726, + 0.020999319851398468, + 0.01335976179689169, + 0.02173866517841816, + 0.01602128893136978, + 0.006359371822327375, + 0.015791194513440132, + 0.010293660685420036, + 0.015852173790335655, + 0.012057468295097351, + 0.0077437893487513065, + 0.007725043222308159, + 0.006461305543780327, + 0.01159240398555994, + 0.013635833747684956, + 0.013337776064872742, + 0.0185526255518198, + 0.009421988390386105, + 0.008140887133777142, + 0.010983478277921677, + 0.009275969117879868, + 0.003502974286675453, + 0.008925321511924267, + 0.004137125331908464, + 0.005977972410619259, + 0.005337142385542393, + 0.004089294001460075, + 0.004181261174380779, + 0.004307842813432217, + 0.003514522220939398, + 0.004201949108392, + 0.004176388960331678, + 0.004020466469228268, + 0.005064988508820534, + 0.00462399423122406, + 0.003635176457464695, + 0.004806476179510355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.08639074862003326 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004111053422093391 + ], + "opposite_logprob_tokens": [ + -5.504111289978027 + ], + "logprob_reward": -0.004111053422093391, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004111053422093391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27522432804107666, + 0.1716299206018448, + 0.06999044120311737, + 0.06859132647514343, + 0.03521248325705528, + 0.12182667851448059, + 0.06859936565160751, + 0.08295081555843353, + 0.05435523763298988, + 0.051619354635477066 + ], + "credit_map": [ + 0.02937004528939724, + 0.04405895620584488, + 0.03672460466623306, + 0.05420059710741043, + 0.023477140814065933, + 0.024364186450839043, + 0.02646896429359913, + 0.036559853702783585, + 0.05879603698849678, + 0.012705862522125244, + 0.06173688545823097, + 0.010935496538877487, + 0.00822404958307743, + 0.005371776409447193, + 0.006993839517235756, + 0.0068659852258861065, + 0.018034402281045914, + 0.005597748328000307, + 0.018483983352780342, + 0.007066196296364069, + 0.00642011733725667, + 0.004549132660031319, + 0.005624909419566393, + 0.004213958978652954, + 0.009232303127646446, + 0.016220979392528534, + 0.00805557332932949, + 0.013400503434240818, + 0.00634124968200922, + 0.004715964198112488, + 0.0060318876057863235, + 0.004592875950038433, + 0.0032001477666199207, + 0.005001551005989313, + 0.00335533800534904, + 0.005023340228945017, + 0.004296425264328718, + 0.0041343411430716515, + 0.005941059906035662, + 0.004260282963514328, + 0.010836737230420113, + 0.026357289403676987, + 0.01523506361991167, + 0.023979781195521355, + 0.012866715900599957, + 0.008334251120686531, + 0.01380810048431158, + 0.010408744215965271, + 0.006268264725804329, + 0.012584278360009193, + 0.007893060334026814, + 0.011304245330393314, + 0.0109626529738307, + 0.006727975327521563, + 0.007263222709298134, + 0.005595674272626638, + 0.008716171607375145, + 0.018594376742839813, + 0.00945692416280508, + 0.016527200117707253, + 0.008848965167999268, + 0.005638950504362583, + 0.008336203172802925, + 0.006832025479525328, + 0.004467401187866926, + 0.009394863620400429, + 0.005367246922105551, + 0.008320227265357971, + 0.008933496661484241, + 0.006101752165704966, + 0.006443039979785681, + 0.005327217746526003, + 0.005422984249889851, + 0.010096754878759384, + 0.006181736011058092, + 0.009702442213892937, + 0.005137593485414982, + 0.0046864645555615425, + 0.005241572391241789, + 0.00514981197193265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.004111053422093391 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746490485966206 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746490485966206, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746490485966206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19621266424655914, + 0.31033122539520264, + 0.05339876562356949, + 0.02308080345392227, + 0.020971691235899925, + 0.03514627739787102, + 0.029041972011327744, + 0.03975379467010498, + 0.09229384362697601, + 0.1997690349817276 + ], + "credit_map": [ + 0.016536563634872437, + 0.021539749577641487, + 0.025975456461310387, + 0.031597476452589035, + 0.03759192302823067, + 0.018981076776981354, + 0.02297838404774666, + 0.02101203426718712, + 0.1078871488571167, + 0.01814166083931923, + 0.11430525034666061, + 0.018319685012102127, + 0.017287593334913254, + 0.011531861498951912, + 0.011028311215341091, + 0.011829691007733345, + 0.012184981256723404, + 0.003419452579692006, + 0.01220713835209608, + 0.004328944254666567, + 0.007243254221975803, + 0.005618189461529255, + 0.004026609472930431, + 0.004370188806205988, + 0.0031643027905374765, + 0.0025997618213295937, + 0.002945861080661416, + 0.003946138545870781, + 0.0031494710128754377, + 0.0023574410006403923, + 0.0028599954675883055, + 0.0020578294061124325, + 0.0020929924212396145, + 0.003377550281584263, + 0.0019325644243508577, + 0.003608676604926586, + 0.0028637356590479612, + 0.0018510015215724707, + 0.003219634061679244, + 0.002025534864515066, + 0.0033767963759601116, + 0.007579921279102564, + 0.004461273085325956, + 0.006882801651954651, + 0.0031419997103512287, + 0.0019982121884822845, + 0.004761443939059973, + 0.0029438247438520193, + 0.002139038173481822, + 0.004956530407071114, + 0.003194277174770832, + 0.00378491822630167, + 0.0049759200774133205, + 0.0035272969398647547, + 0.003334400011226535, + 0.003129588672891259, + 0.005161089822649956, + 0.009066036902368069, + 0.005239544436335564, + 0.008515395224094391, + 0.003609251696616411, + 0.0022210602182894945, + 0.0035297314170747995, + 0.0024116775020956993, + 0.00812030490487814, + 0.010903097689151764, + 0.008644942194223404, + 0.012075738981366158, + 0.015769170597195625, + 0.011764341965317726, + 0.013395584188401699, + 0.011620663106441498, + 0.01888686791062355, + 0.020735349506139755, + 0.019582314416766167, + 0.02523888647556305, + 0.03449214622378349, + 0.028737595304846764, + 0.02855972945690155, + 0.02353615127503872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.006746490485966206 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630369160324335 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630369160324335, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630369160324335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2521313726902008, + 0.18819928169250488, + 0.09218568354845047, + 0.034988291561603546, + 0.05246614292263985, + 0.07128308713436127, + 0.08120369911193848, + 0.09298238903284073, + 0.04260125756263733, + 0.09195876121520996 + ], + "credit_map": [ + 0.030519558116793633, + 0.032117560505867004, + 0.033020783215761185, + 0.041304100304841995, + 0.030400998890399933, + 0.025077391415834427, + 0.02484392561018467, + 0.03484706953167915, + 0.039422839879989624, + 0.012938951142132282, + 0.03982064127922058, + 0.019546901807188988, + 0.022119315341114998, + 0.019708829000592232, + 0.01602276600897312, + 0.018619032576680183, + 0.010157277807593346, + 0.005924127995967865, + 0.010484295897185802, + 0.009170472621917725, + 0.017554763704538345, + 0.01300568226724863, + 0.013162674382328987, + 0.01272638700902462, + 0.004684635903686285, + 0.004569981712847948, + 0.0040471781976521015, + 0.005871628411114216, + 0.004549087956547737, + 0.0032801013439893723, + 0.00387882674112916, + 0.004106850363314152, + 0.0062576718628406525, + 0.010343760251998901, + 0.00596777955070138, + 0.00902995839715004, + 0.005076117347925901, + 0.004585118032991886, + 0.005284769460558891, + 0.005920971278101206, + 0.0064270151779055595, + 0.011851604096591473, + 0.00756645854562521, + 0.009697708301246166, + 0.010943162254989147, + 0.006526743061840534, + 0.009917627088725567, + 0.008352767676115036, + 0.005950188264250755, + 0.0151206711307168, + 0.009879733435809612, + 0.013568551279604435, + 0.012380839325487614, + 0.008512025699019432, + 0.00854913704097271, + 0.007242555730044842, + 0.010484348051249981, + 0.016730275005102158, + 0.011108196340501308, + 0.017376242205500603, + 0.010668503120541573, + 0.006799750030040741, + 0.010790669359266758, + 0.009024403057992458, + 0.00338909775018692, + 0.006360958330333233, + 0.004421690013259649, + 0.005420852452516556, + 0.007137575652450323, + 0.0057813869789242744, + 0.005855460185557604, + 0.00423423619940877, + 0.007648811209946871, + 0.012136470526456833, + 0.008650460280478, + 0.011584625579416752, + 0.014424366876482964, + 0.013463489711284637, + 0.010870973579585552, + 0.013179564848542213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.007630369160324335 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633446827530861 + ], + "opposite_logprob_tokens": [ + -3.651334524154663 + ], + "logprob_reward": -0.02633446827530861, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633446827530861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16878309845924377, + 0.05381588265299797, + 0.13210546970367432, + 0.10108162462711334, + 0.048345230519771576, + 0.0230008102953434, + 0.10582196712493896, + 0.12059662491083145, + 0.11921589821577072, + 0.12723340094089508 + ], + "credit_map": [ + 0.016386572271585464, + 0.03583545237779617, + 0.01894838735461235, + 0.04308699071407318, + 0.013679132796823978, + 0.010854140855371952, + 0.015322118066251278, + 0.014670325443148613, + 0.009333203546702862, + 0.007168657146394253, + 0.009782649576663971, + 0.007536852266639471, + 0.006569138262420893, + 0.004693150520324707, + 0.004455626476556063, + 0.004276607185602188, + 0.03601079061627388, + 0.013411863707005978, + 0.03644084185361862, + 0.014318509958684444, + 0.008500974625349045, + 0.006480167154222727, + 0.009116634726524353, + 0.007825683802366257, + 0.0116883534938097, + 0.022051764652132988, + 0.014013647101819515, + 0.021205153316259384, + 0.0076894904486835, + 0.007692388258874416, + 0.008408668451011181, + 0.008332159370183945, + 0.0060227904468774796, + 0.00960910227149725, + 0.006325670052319765, + 0.008859173394739628, + 0.005479985848069191, + 0.0033596318680793047, + 0.0055891722440719604, + 0.0030997106805443764, + 0.002362842671573162, + 0.004829930141568184, + 0.002652761060744524, + 0.0035934506449848413, + 0.0025267063174396753, + 0.0021971154492348433, + 0.002445991151034832, + 0.00239201495423913, + 0.00946081429719925, + 0.02226620726287365, + 0.013769690878689289, + 0.019373741000890732, + 0.012307437136769295, + 0.007386850193142891, + 0.010936351493000984, + 0.01032087579369545, + 0.012476337142288685, + 0.030469659715890884, + 0.013743931427598, + 0.024696871638298035, + 0.010351677425205708, + 0.0068186367861926556, + 0.011208753101527691, + 0.010830759070813656, + 0.011939498595893383, + 0.03571063280105591, + 0.012525410391390324, + 0.0205440204590559, + 0.011920081451535225, + 0.008510113693773746, + 0.00857063289731741, + 0.009495503269135952, + 0.015432077459990978, + 0.03883766382932663, + 0.017419788986444473, + 0.021828068420290947, + 0.009782187640666962, + 0.007667802274227142, + 0.009641975164413452, + 0.0066238464787602425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.02633446827530861 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "locked, completely still long shot gazes steadily at a nearly vertical, weathered, gray rock face that dominates the frame. To one side, a massive waterfall cascades down from the heights like a galaxy, crashing into the deep pool below, stirring up a cloud of dancing mist. This dense mist envelops the lower half of the scene, slowly curling and rising in front of the cliff. Cerebral light filters through thick clouds and pervasive vapor, transforming into a soft, evenly diffused glow, gently illuminating every detail of the rock face. On this natural, monolithic canvas, the majestic figures of the Four Heavenly Kings are carved in massive relief. Clad in intricate armor and wielding their iconic magical instruments, their faces stern and resolute, their eyes wide open, their expressions solemn and majestic. The surface of the statue is not smooth stone, but rather exhibits the rich, granular texture characteristic of rock painting. Its vibrant and quaint colors, richly applied in cinnabar red, azurite, malachite green, and ochre, appear exceptionally saturated and profound in the humid air. Countless tiny water droplets cling to the statue's mottled hues, shimmering and adding a touch of vividness to the otherwise tranquil scene. The entire painting is imbued with a sense of grandeur, mystery, and sacred power, creating an awe-inspiring atmosphere of tranquility.", + "seed": 6783, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_8/index_3182_seed_6783_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10340841822326183, + "vlm_logprob": -0.10340841822326183, + "vlm_yes_no_margin": 2.7249999761581423, + "vlm_token_credit": -0.10340841822326183, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2753854990005493 + ], + "opposite_logprob_tokens": [ + -1.5253854990005493 + ], + "logprob_reward": -0.2753854990005493, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2753854990005493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38710102438926697, + 0.09526760876178741, + 0.17862145602703094, + 0.0738486498594284, + 0.04658475145697594, + 0.031537383794784546, + 0.015790073201060295, + 0.07907938212156296, + 0.016771303489804268, + 0.07539836317300797 + ], + "credit_map": [ + 0.06743653118610382, + 0.05563078075647354, + 0.05132400989532471, + 0.046577781438827515, + 0.0476059690117836, + 0.04572348669171333, + 0.03660315275192261, + 0.036199335008859634, + 0.01409673411399126, + 0.025272343307733536, + 0.01353888213634491, + 0.015356077812612057, + 0.009257876314222813, + 0.00586456386372447, + 0.006082493811845779, + 0.005798647180199623, + 0.022237194702029228, + 0.017678020521998405, + 0.03981618955731392, + 0.048402391374111176, + 0.014050000347197056, + 0.008861221373081207, + 0.013593648560345173, + 0.01398278959095478, + 0.013597811572253704, + 0.006763572338968515, + 0.012025385163724422, + 0.012526331469416618, + 0.007135339546948671, + 0.006251949816942215, + 0.007676859386265278, + 0.00787140242755413, + 0.005748317576944828, + 0.005766641348600388, + 0.00716786365956068, + 0.00732728186994791, + 0.005605638958513737, + 0.004116591066122055, + 0.005731421522796154, + 0.0051209949888288975, + 0.005366826895624399, + 0.002985924482345581, + 0.006302147172391415, + 0.0059324572794139385, + 0.002620491897687316, + 0.0023711882531642914, + 0.0040088738314807415, + 0.0019494753796607256, + 0.0029238243587315083, + 0.002059060847386718, + 0.002590335439890623, + 0.002067716559395194, + 0.0015045494074001908, + 0.0013519574422389269, + 0.001999150961637497, + 0.0012934805126860738, + 0.011394260451197624, + 0.008959680795669556, + 0.009701131843030453, + 0.01293142605572939, + 0.00970095582306385, + 0.006019902881234884, + 0.013387463055551052, + 0.006984565407037735, + 0.0030387742444872856, + 0.0019891024567186832, + 0.0022511989809572697, + 0.0020425443071871996, + 0.0018601745832711458, + 0.0017395007889717817, + 0.002296116901561618, + 0.001553892157971859, + 0.014004017226397991, + 0.007824167609214783, + 0.010007817298173904, + 0.009861761704087257, + 0.008624397218227386, + 0.008070890791714191, + 0.0104133952409029, + 0.006591918412595987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.2753854990005493 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283187687397003 + ], + "opposite_logprob_tokens": [ + -2.017831802368164 + ], + "logprob_reward": -0.14283187687397003, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14283187687397003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4388181269168854, + 0.19105957448482513, + 0.049617912620306015, + 0.0354723297059536, + 0.03356993943452835, + 0.020747477188706398, + 0.021534306928515434, + 0.11709757149219513, + 0.026678213849663734, + 0.06540454924106598 + ], + "credit_map": [ + 0.08558932691812515, + 0.07211552560329437, + 0.06291594356298447, + 0.05441242828965187, + 0.04950707033276558, + 0.0444747693836689, + 0.03627479821443558, + 0.03352825343608856, + 0.030150720849633217, + 0.05715947225689888, + 0.027487896382808685, + 0.03538781404495239, + 0.014897523447871208, + 0.008219428360462189, + 0.010002994909882545, + 0.007753719110041857, + 0.006680293008685112, + 0.004306945484131575, + 0.009135914966464043, + 0.009904910810291767, + 0.005909886211156845, + 0.0044879307970404625, + 0.004635412245988846, + 0.004556616768240929, + 0.007328841369599104, + 0.0038151575718075037, + 0.00520136347040534, + 0.005285424180328846, + 0.004273928701877594, + 0.0032845025416463614, + 0.003555480157956481, + 0.0027276326436549425, + 0.005475906189531088, + 0.004046278074383736, + 0.0042487881146371365, + 0.006366862915456295, + 0.004146868828684092, + 0.0027890941128134727, + 0.0039145988412201405, + 0.0025815411936491728, + 0.0041745188646018505, + 0.0023007881827652454, + 0.003253093920648098, + 0.002830429933965206, + 0.0023481284733861685, + 0.002070985035970807, + 0.0020301365293562412, + 0.0017393957823514938, + 0.0044758995063602924, + 0.0024441045243293047, + 0.003717380575835705, + 0.0029995932709425688, + 0.0019428841769695282, + 0.0019679733086377382, + 0.002172369509935379, + 0.0018141025211662054, + 0.02314501255750656, + 0.012903681956231594, + 0.017248867079615593, + 0.01767265424132347, + 0.011795193888247013, + 0.008280203677713871, + 0.01752549596130848, + 0.008526467718183994, + 0.006265673320740461, + 0.003272783476859331, + 0.005904807709157467, + 0.003991992212831974, + 0.0013835065765306354, + 0.0015415316447615623, + 0.0021115229465067387, + 0.00220639375038445, + 0.01219793502241373, + 0.007782976143062115, + 0.009546712040901184, + 0.01044389232993126, + 0.006521699484437704, + 0.005882084369659424, + 0.008494601584970951, + 0.004534644540399313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.14283187687397003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496761202812195 + ], + "opposite_logprob_tokens": [ + -2.9299676418304443 + ], + "logprob_reward": -0.05496761202812195, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496761202812195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4847120940685272, + 0.19031943380832672, + 0.053637582808732986, + 0.029151074588298798, + 0.03426288068294525, + 0.022666584700345993, + 0.027088504284620285, + 0.07265623658895493, + 0.021395239979028702, + 0.06411034613847733 + ], + "credit_map": [ + 0.08737838268280029, + 0.07065650075674057, + 0.06635720282793045, + 0.05755406245589256, + 0.058898068964481354, + 0.054065100848674774, + 0.04496998339891434, + 0.04483276605606079, + 0.025611743330955505, + 0.06283053755760193, + 0.031749822199344635, + 0.03245314210653305, + 0.013760868459939957, + 0.007999557070434093, + 0.008763779886066914, + 0.007149986457079649, + 0.007615641690790653, + 0.004796512424945831, + 0.009776259772479534, + 0.01048183161765337, + 0.006458004005253315, + 0.0047526611015200615, + 0.004851715173572302, + 0.004904960747808218, + 0.004275350831449032, + 0.0033092505764216185, + 0.0030987649224698544, + 0.004212123807519674, + 0.004602537024766207, + 0.0029105425346642733, + 0.003911813721060753, + 0.0028306907042860985, + 0.005482540000230074, + 0.004239742644131184, + 0.004128283355385065, + 0.00663737440481782, + 0.0042525120079517365, + 0.002833142876625061, + 0.004217740148305893, + 0.0024715436156839132, + 0.004605432972311974, + 0.0021040630526840687, + 0.0031989042181521654, + 0.0032021375373005867, + 0.0024891558568924665, + 0.0022636759094893932, + 0.0022385232150554657, + 0.002564690774306655, + 0.005099679343402386, + 0.0030554314143955708, + 0.0036107541527599096, + 0.004041230771690607, + 0.002898033009842038, + 0.00235718279145658, + 0.0035122213885188103, + 0.0025139700155705214, + 0.013111632317304611, + 0.008268420584499836, + 0.009997610934078693, + 0.010838277637958527, + 0.007787943352013826, + 0.005633931141346693, + 0.01127647515386343, + 0.0057419417425990105, + 0.004205671139061451, + 0.002710541943088174, + 0.0039939009584486485, + 0.0037845629267394543, + 0.001348756835795939, + 0.00137713726144284, + 0.0020116453524678946, + 0.0019630242604762316, + 0.012398109771311283, + 0.007204046472907066, + 0.008091623894870281, + 0.008829253725707531, + 0.0068458388559520245, + 0.006530458107590675, + 0.008732065558433533, + 0.005478948354721069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.05496761202812195 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581956952810287 + ], + "opposite_logprob_tokens": [ + -3.8955819606781006 + ], + "logprob_reward": -0.020581956952810287, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581956952810287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48469316959381104, + 0.11699037253856659, + 0.0674784854054451, + 0.045638661831617355, + 0.06062925606966019, + 0.029520342126488686, + 0.02680174820125103, + 0.10056093335151672, + 0.026652945205569267, + 0.04103405401110649 + ], + "credit_map": [ + 0.08776546269655228, + 0.06260353326797485, + 0.06104644387960434, + 0.05762207880616188, + 0.0582149438560009, + 0.05968468636274338, + 0.047648731619119644, + 0.05010733753442764, + 0.02110530622303486, + 0.028774002566933632, + 0.011934250593185425, + 0.01870143599808216, + 0.011048052459955215, + 0.007793174125254154, + 0.009187962859869003, + 0.008446184918284416, + 0.009189439006149769, + 0.005489862058311701, + 0.011262447573244572, + 0.012431038543581963, + 0.00849910918623209, + 0.006615408696234226, + 0.006354172248393297, + 0.007637007161974907, + 0.0059560444205999374, + 0.004615091253072023, + 0.004991339985281229, + 0.0076266685500741005, + 0.007264274172484875, + 0.004610028583556414, + 0.005807030946016312, + 0.004768184386193752, + 0.009685439057648182, + 0.008381722494959831, + 0.006222181022167206, + 0.009060101583600044, + 0.008061018772423267, + 0.005513423588126898, + 0.007650250568985939, + 0.0060551222413778305, + 0.0056345402263104916, + 0.0031567171681672335, + 0.0035372921265661716, + 0.0037084391806274652, + 0.0034634205512702465, + 0.003501046681776643, + 0.0030898789409548044, + 0.003429008647799492, + 0.004176369868218899, + 0.0028988104313611984, + 0.0032215528190135956, + 0.0036164643242955208, + 0.002987447427585721, + 0.0029944581910967827, + 0.003456977428868413, + 0.003449666080996394, + 0.015608049929141998, + 0.011294410564005375, + 0.01358655747026205, + 0.014830715954303741, + 0.011542489752173424, + 0.007908955216407776, + 0.01591845043003559, + 0.009871300309896469, + 0.004710891284048557, + 0.003968437202274799, + 0.003701835172250867, + 0.005832675378769636, + 0.0015939339064061642, + 0.0015295122284442186, + 0.00298378081060946, + 0.0023318789899349213, + 0.006255243439227343, + 0.00466323783621192, + 0.006499574985355139, + 0.005603889934718609, + 0.004096700809895992, + 0.0040444498881697655, + 0.005926971789449453, + 0.003943986259400845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.020581956952810287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot locked, completely still, and framed as a long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a nearly vertical, weathered, gray rock face dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the Four Heavenly Kings carved in massive relief on the rock face, wearing intricate armor and wielding magical instruments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the statues exhibit a granular rock painting texture with vibrant colors including cinnabar red, azurite, malachite green, and ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a massive waterfall cascading down the side of the cliff into a deep pool below, surrounded by dense mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023275146260857582 + ], + "opposite_logprob_tokens": [ + -3.773275136947632 + ], + "logprob_reward": -0.023275146260857582, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023275146260857582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4724231958389282, + 0.19082391262054443, + 0.030587321147322655, + 0.03630668297410011, + 0.03634829446673393, + 0.021671151742339134, + 0.03332211449742317, + 0.0942150130867958, + 0.020460093393921852, + 0.06384222954511642 + ], + "credit_map": [ + 0.09546664357185364, + 0.06984435766935349, + 0.06350968033075333, + 0.05486202612519264, + 0.05450607091188431, + 0.04924686253070831, + 0.04191608354449272, + 0.04307148978114128, + 0.027215223759412766, + 0.06171143054962158, + 0.022285448387265205, + 0.04104571416974068, + 0.013372262939810753, + 0.008078558370471, + 0.00891347136348486, + 0.008201799355447292, + 0.0036933389492332935, + 0.002535347593948245, + 0.0034482572227716446, + 0.004781806375831366, + 0.004962884355336428, + 0.0035914629697799683, + 0.003938126377761364, + 0.003636098001152277, + 0.007669266778975725, + 0.0036828198935836554, + 0.006916398648172617, + 0.005271356552839279, + 0.004145773593336344, + 0.0030461715068668127, + 0.0032246937043964863, + 0.0023502043914049864, + 0.0043371631763875484, + 0.004329691641032696, + 0.00464740302413702, + 0.005991344340145588, + 0.0053577604703605175, + 0.003071647137403488, + 0.005013859365135431, + 0.003599424846470356, + 0.0038305011112242937, + 0.0027809953317046165, + 0.0029900320805609226, + 0.003073085332289338, + 0.0023856735788285732, + 0.0020813627634197474, + 0.0030032042413949966, + 0.0015262967208400369, + 0.005222473759204149, + 0.004952233750373125, + 0.004301617853343487, + 0.005515731871128082, + 0.003202325664460659, + 0.002817069413140416, + 0.0043235355988144875, + 0.0029871254228055477, + 0.016862472519278526, + 0.010929297655820847, + 0.013841772451996803, + 0.014317546971142292, + 0.009803137741982937, + 0.006875226274132729, + 0.014507454819977283, + 0.007078102789819241, + 0.0030070242937654257, + 0.003022279590368271, + 0.0033277624752372503, + 0.0034939725883305073, + 0.0018298292998224497, + 0.001694952603429556, + 0.0023056070785969496, + 0.0017786647658795118, + 0.010934781283140182, + 0.008951555006206036, + 0.010383717715740204, + 0.010968868620693684, + 0.005579570308327675, + 0.005473711062222719, + 0.0074231927283108234, + 0.004126827698200941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.023275146260857582 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_9/index_tensor([1992, 1992])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_9/index_tensor([1992, 1992])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..01760b4fae03b405dae19d74b74cfa8f1bbebc54 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_9/index_tensor([1992, 1992])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "still, vintage film-like close-up captures the face of a majestic Siberian tiger with unwavering precision. In the pale, soft, diffuse light of a winter afternoon, its dense, orange-white fur is sharply outlined by deep black stripes. Its deep, amber eyes exude a primal alertness and majesty, while a few white whiskers, stained with tiny ice crystals, quiver softly on either side of its damp, black nose. The snow-capped coniferous forest in the background is completely blurred by the shallow depth of field, with only the faint outlines of the trees visible. The tiger exhales, solemnly and calmly, slowly and deeply. The warm air instantly condenses into a visible, swirling white mist in the chill air, which is then gently blown away by the cold wind. The entire picture has obvious film grain and the color saturation is slightly reduced, creating a quiet, solemn and wild and poetic atmosphere. In the end, this shot full of power and life slowly fades out to pure black.", + "seed": 28809, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_9/index_1992_seed_28809_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14243947621434927, + "vlm_logprob": -0.14243947621434927, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.14243947621434927, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007475498132407665 + ], + "opposite_logprob_tokens": [ + -6.50747537612915 + ], + "logprob_reward": -0.007475498132407665, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.007475498132407665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05970584228634834, + 0.016751667484641075, + 0.0261521153151989, + 0.03164702653884888, + 0.04244785010814667, + 0.025550970807671547, + 0.06328269094228745, + 0.04165629670023918, + 0.14618290960788727, + 0.5466225743293762 + ], + "credit_map": [ + 0.006375026423484087, + 0.0062967343255877495, + 0.010981339029967785, + 0.006003598216921091, + 0.006154618225991726, + 0.01011824980378151, + 0.006837598979473114, + 0.00693867914378643, + 0.0019434818532317877, + 0.0016302983276546001, + 0.002819450804963708, + 0.0016750060021877289, + 0.0014881219249218702, + 0.0029185605235397816, + 0.0024176365695893764, + 0.0018591121770441532, + 0.0033176562283188105, + 0.0020210661459714174, + 0.00414395984262228, + 0.001867187093012035, + 0.0027729386929422617, + 0.006041348911821842, + 0.0026216355618089437, + 0.0033663217909634113, + 0.003745126537978649, + 0.0032629098277539015, + 0.006136946380138397, + 0.0023211475927382708, + 0.0025355126708745956, + 0.006653545890003443, + 0.002829616889357567, + 0.004162217490375042, + 0.004023212939500809, + 0.00345404795370996, + 0.007701063062995672, + 0.004356550518423319, + 0.004826863296329975, + 0.007006623782217503, + 0.006445226725190878, + 0.004634261596947908, + 0.0025755646638572216, + 0.003107575001195073, + 0.0030421412084251642, + 0.0027607891242951155, + 0.0033437409438192844, + 0.004101835191249847, + 0.003965935204178095, + 0.002653389936313033, + 0.006261513568460941, + 0.004646657966077328, + 0.008184648118913174, + 0.005682367831468582, + 0.00674362201243639, + 0.015792138874530792, + 0.008569197729229927, + 0.0074025415815413, + 0.005250918213278055, + 0.0036885207518935204, + 0.006506321020424366, + 0.0039589968509972095, + 0.003347645280882716, + 0.009900994598865509, + 0.0042833928018808365, + 0.004719507414847612, + 0.014680718071758747, + 0.01624906063079834, + 0.019557954743504524, + 0.019478633999824524, + 0.01789739541709423, + 0.02674097940325737, + 0.016300028190016747, + 0.015278136357665062, + 0.03755835443735123, + 0.06884826719760895, + 0.05607398971915245, + 0.07624346762895584, + 0.05804569274187088, + 0.09620220959186554, + 0.0703376978635788, + 0.08331288397312164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.007475498132407665 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255748514086008 + ], + "opposite_logprob_tokens": [ + -4.387557506561279 + ], + "logprob_reward": -0.01255748514086008, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255748514086008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041331060230731964, + 0.024484358727931976, + 0.03786158561706543, + 0.032972294837236404, + 0.032292645424604416, + 0.042438823729753494, + 0.0539335198700428, + 0.07861951738595963, + 0.22019144892692566, + 0.4358747899532318 + ], + "credit_map": [ + 0.0051208301447331905, + 0.004743787460029125, + 0.007072700187563896, + 0.00377999828197062, + 0.004401996731758118, + 0.0065937708131968975, + 0.00456309225410223, + 0.00505488459020853, + 0.002727871062234044, + 0.0020026916172355413, + 0.004950241185724735, + 0.002164611592888832, + 0.0023457854986190796, + 0.00503586558625102, + 0.002298870123922825, + 0.002958420431241393, + 0.004628367722034454, + 0.004865332972258329, + 0.005818055476993322, + 0.003714928636327386, + 0.004306817892938852, + 0.006120257079601288, + 0.004498335998505354, + 0.003909487742930651, + 0.0041169882752001286, + 0.003521763952448964, + 0.0038902508094906807, + 0.0030328910797834396, + 0.0041750469245016575, + 0.00549778388813138, + 0.004385981243103743, + 0.0043515912257134914, + 0.0038538717199116945, + 0.003359749913215637, + 0.004524650517851114, + 0.0032190512865781784, + 0.0037384098395705223, + 0.0058382051065564156, + 0.003899704897776246, + 0.0038590040057897568, + 0.0040024276822805405, + 0.003501736791804433, + 0.006375472526997328, + 0.0032608003821223974, + 0.0040929135866463184, + 0.009174846112728119, + 0.007098379544913769, + 0.00493224710226059, + 0.006154063623398542, + 0.003278750227764249, + 0.0071645332500338554, + 0.003996653016656637, + 0.005122689995914698, + 0.014847184531390667, + 0.00695620896294713, + 0.006413430906832218, + 0.00724519370123744, + 0.00678236735984683, + 0.012754335068166256, + 0.008037388324737549, + 0.006747707258909941, + 0.019672615453600883, + 0.008470003493130207, + 0.008909904398024082, + 0.01976466365158558, + 0.01580462045967579, + 0.02409767359495163, + 0.01788342371582985, + 0.01922396384179592, + 0.07104385644197464, + 0.021919570863246918, + 0.030453680083155632, + 0.0435602180659771, + 0.04631636291742325, + 0.08290670812129974, + 0.06792324781417847, + 0.04578762128949165, + 0.03945046290755272, + 0.06528304517269135, + 0.04464716091752052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.01255748514086008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083937250077724 + ], + "opposite_logprob_tokens": [ + -4.511084079742432 + ], + "logprob_reward": -0.011083937250077724, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083937250077724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045092880725860596, + 0.07568724453449249, + 0.056232333183288574, + 0.07120851427316666, + 0.04281097650527954, + 0.04074186459183693, + 0.029743418097496033, + 0.06049659475684166, + 0.22752974927425385, + 0.35045647621154785 + ], + "credit_map": [ + 0.004374381620436907, + 0.004600451327860355, + 0.008688699454069138, + 0.004527138080447912, + 0.004339807666838169, + 0.007450849749147892, + 0.005728899035602808, + 0.005382656119763851, + 0.010006224736571312, + 0.005822190083563328, + 0.013724521733820438, + 0.0065554105676710606, + 0.00532762985676527, + 0.016341082751750946, + 0.007684929296374321, + 0.010225256904959679, + 0.007115071639418602, + 0.004907398950308561, + 0.009872808121144772, + 0.004415187519043684, + 0.0057750968262553215, + 0.011935286223888397, + 0.005350586958229542, + 0.006860898341983557, + 0.01099640503525734, + 0.005130084231495857, + 0.01141638495028019, + 0.004994708113372326, + 0.005964416544884443, + 0.014910133555531502, + 0.006277556996792555, + 0.011518828570842743, + 0.004947197623550892, + 0.0033628621604293585, + 0.007227362133562565, + 0.004650205373764038, + 0.003935549408197403, + 0.006673756521195173, + 0.007136237807571888, + 0.004877805709838867, + 0.003825624007731676, + 0.004369170870631933, + 0.005396798253059387, + 0.004389413632452488, + 0.004293164703994989, + 0.009001175872981548, + 0.005493769887834787, + 0.003972745966166258, + 0.002835263963788748, + 0.002239339053630829, + 0.004732701927423477, + 0.0024324802216142416, + 0.003478514263406396, + 0.0063984883017838, + 0.004504922777414322, + 0.0031217080540955067, + 0.0062184990383684635, + 0.004483369644731283, + 0.011632910929620266, + 0.004606742411851883, + 0.006254660896956921, + 0.01392403244972229, + 0.006485693622380495, + 0.006890688091516495, + 0.01931249350309372, + 0.01274929754436016, + 0.025807876139879227, + 0.01814933866262436, + 0.021683499217033386, + 0.07945455610752106, + 0.01779717579483986, + 0.032575495541095734, + 0.03447427228093147, + 0.038912560790777206, + 0.05387573316693306, + 0.04468848928809166, + 0.048582471907138824, + 0.04477867856621742, + 0.05023573338985443, + 0.034908536821603775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.011083937250077724 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215697288513184 + ], + "opposite_logprob_tokens": [ + -1.5021569728851318 + ], + "logprob_reward": -0.25215697288513184, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215697288513184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04155311733484268, + 0.03292617201805115, + 0.02925071306526661, + 0.053408198058605194, + 0.01969565637409687, + 0.03761335089802742, + 0.05079331248998642, + 0.08548668026924133, + 0.12369053065776825, + 0.5255823135375977 + ], + "credit_map": [ + 0.005548674613237381, + 0.003625900484621525, + 0.006363309919834137, + 0.0038370073307305574, + 0.00344035099260509, + 0.009716189466416836, + 0.003710696939378977, + 0.005310986191034317, + 0.003543211380019784, + 0.002707665553316474, + 0.006773742381483316, + 0.0028013449627906084, + 0.0034488143865019083, + 0.005448944866657257, + 0.00322764297015965, + 0.00497480621561408, + 0.0031772744841873646, + 0.0023104315623641014, + 0.007051384076476097, + 0.0024140800815075636, + 0.003994681406766176, + 0.003799059893935919, + 0.002943776547908783, + 0.003560022683814168, + 0.007699818350374699, + 0.0037826593033969402, + 0.007830996997654438, + 0.003792655421420932, + 0.0049615525640547276, + 0.011488142423331738, + 0.004801375791430473, + 0.009050995111465454, + 0.001970212673768401, + 0.0017152759246528149, + 0.002763979136943817, + 0.0021288450807332993, + 0.0021421003621071577, + 0.002915422199293971, + 0.0032756926957517862, + 0.0027841273695230484, + 0.0034704844001680613, + 0.003783784806728363, + 0.004962977487593889, + 0.003392895683646202, + 0.003721721237525344, + 0.008543611504137516, + 0.0060409740544855595, + 0.003696902422234416, + 0.006472448352724314, + 0.00315163959749043, + 0.005342813208699226, + 0.003557446412742138, + 0.0038350368849933147, + 0.016075672581791878, + 0.006616263184696436, + 0.005741993896663189, + 0.008864029310643673, + 0.005200905259698629, + 0.009876834228634834, + 0.005602851044386625, + 0.006990005262196064, + 0.028380539268255234, + 0.008766665123403072, + 0.011804849840700626, + 0.012547902762889862, + 0.01003567036241293, + 0.018209559842944145, + 0.011606580577790737, + 0.010630904696881771, + 0.03168611600995064, + 0.012747432105243206, + 0.016226356849074364, + 0.04464448243379593, + 0.055073171854019165, + 0.12427689135074615, + 0.07821539044380188, + 0.05027689039707184, + 0.04530513286590576, + 0.08030963689088821, + 0.047480709850788116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.25215697288513184 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42892348766326904 + ], + "opposite_logprob_tokens": [ + -1.053923487663269 + ], + "logprob_reward": -0.42892348766326904, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42892348766326904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038057148456573486, + 0.014626896008849144, + 0.019235920161008835, + 0.009134935215115547, + 0.035604000091552734, + 0.023062774911522865, + 0.032644305378198624, + 0.039938461035490036, + 0.10989266633987427, + 0.6778028607368469 + ], + "credit_map": [ + 0.004411684814840555, + 0.004435465205460787, + 0.006024200469255447, + 0.003307655220851302, + 0.0039007216691970825, + 0.006253882311284542, + 0.004252251237630844, + 0.005471285432577133, + 0.0015180596383288503, + 0.0015993161359801888, + 0.0019299156265333295, + 0.0012339861132204533, + 0.0018796146614477038, + 0.0029056842904537916, + 0.0016370905796065927, + 0.0019232293125241995, + 0.0022706615272909403, + 0.001510400790721178, + 0.003661096328869462, + 0.001470727613195777, + 0.002554738661274314, + 0.003468630136922002, + 0.0019285466987639666, + 0.0023711177054792643, + 0.0009661117219366133, + 0.0010760925943031907, + 0.0013300372520461679, + 0.0008960295817814767, + 0.0009177583269774914, + 0.0015614167787134647, + 0.0008082177955657244, + 0.0015792715130373836, + 0.004435537848621607, + 0.0022872304543852806, + 0.004540603142231703, + 0.0027007916942238808, + 0.0030109602957963943, + 0.008944886736571789, + 0.005517063196748495, + 0.004166929051280022, + 0.002564892638474703, + 0.001559051568619907, + 0.0027621600311249495, + 0.001963607268407941, + 0.0021008478943258524, + 0.004559644963592291, + 0.00496238749474287, + 0.0025901843328028917, + 0.0034455701243132353, + 0.0023367442190647125, + 0.00527408579364419, + 0.0029309347737580538, + 0.0035782225895673037, + 0.006917692720890045, + 0.004456051625311375, + 0.0037050044629722834, + 0.0035734418779611588, + 0.004462673328816891, + 0.007037593517452478, + 0.004639117978513241, + 0.004545858129858971, + 0.007018140517175198, + 0.0046091144904494286, + 0.0040525211952626705, + 0.00966174528002739, + 0.00855475477874279, + 0.010971860028803349, + 0.010740174911916256, + 0.01098653580993414, + 0.03365485742688179, + 0.01116653997451067, + 0.01415619719773531, + 0.06434435397386551, + 0.06084107980132103, + 0.06927905231714249, + 0.07812273502349854, + 0.0883680209517479, + 0.14561432600021362, + 0.08480525016784668, + 0.08642802387475967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.42892348766326904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, vintage film-like close-up captures the face of a majestic Siberian tiger with unwavering precision. In the pale, soft, diffuse light of a winter afternoon, its dense, orange-white fur is sharply outlined by deep black stripes. Its deep, amber eyes exude a primal alertness and majesty, while a few white whiskers, stained with tiny ice crystals, quiver softly on either side of its damp, black nose. The snow-capped coniferous forest in the background is completely blurred by the shallow depth of field, with only the faint outlines of the trees visible. The tiger exhales, solemnly and calmly, slowly and deeply. The warm air instantly condenses into a visible, swirling white mist in the chill air, which is then gently blown away by the cold wind. The entire picture has obvious film grain and the color saturation is slightly reduced, creating a quiet, solemn and wild and poetic atmosphere. In the end, this shot full of power and life slowly fades out to pure black.", + "seed": 28810, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_9/index_1992_seed_28810_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12708477964624762, + "vlm_logprob": -0.12708477964624762, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.12708477964624762, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009148347191512585 + ], + "opposite_logprob_tokens": [ + -5.759148120880127 + ], + "logprob_reward": -0.009148347191512585, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.009148347191512585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06674465537071228, + 0.06994619220495224, + 0.15849389135837555, + 0.0842108279466629, + 0.057746611535549164, + 0.07709436118602753, + 0.04707472771406174, + 0.17528463900089264, + 0.13850553333759308, + 0.12489853799343109 + ], + "credit_map": [ + 0.009450778365135193, + 0.007803427986800671, + 0.014511045999825, + 0.00857899896800518, + 0.006998008117079735, + 0.005510231945663691, + 0.008276944980025291, + 0.005615215748548508, + 0.011484182439744473, + 0.009314438328146935, + 0.006885522976517677, + 0.009296437725424767, + 0.008648593910038471, + 0.006026722956448793, + 0.012019655667245388, + 0.006270634010434151, + 0.018024640157818794, + 0.022194096818566322, + 0.021539080888032913, + 0.023063575848937035, + 0.025652270764112473, + 0.011396772228181362, + 0.02270592749118805, + 0.013917531818151474, + 0.01425210852175951, + 0.008467120118439198, + 0.009930259548127651, + 0.010053824633359909, + 0.008249261416494846, + 0.01046661101281643, + 0.009650610387325287, + 0.013141031377017498, + 0.006086393725126982, + 0.00876343809068203, + 0.009144252166152, + 0.009244921617209911, + 0.007348921615630388, + 0.004456030670553446, + 0.007054693531244993, + 0.005647958256304264, + 0.008810724131762981, + 0.008082673884928226, + 0.018699467182159424, + 0.010189476422965527, + 0.00799570046365261, + 0.005681677721440792, + 0.008985009044408798, + 0.008649636059999466, + 0.005187402945011854, + 0.008706251159310341, + 0.006515115965157747, + 0.006526147481054068, + 0.006480270065367222, + 0.002438888419419527, + 0.006797778885811567, + 0.004422870464622974, + 0.019468316808342934, + 0.029635731130838394, + 0.020455816760659218, + 0.022977309301495552, + 0.02021368034183979, + 0.01714996062219143, + 0.02073524333536625, + 0.024648571386933327, + 0.009250769391655922, + 0.012842237018048763, + 0.007359567563980818, + 0.010929277166724205, + 0.012507587671279907, + 0.039695534855127335, + 0.010197325609624386, + 0.0357232429087162, + 0.008150782436132431, + 0.013014194555580616, + 0.007366916164755821, + 0.012047032825648785, + 0.007335108704864979, + 0.03378394618630409, + 0.007926277816295624, + 0.03527426719665527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.009148347191512585 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751582026481628 + ], + "opposite_logprob_tokens": [ + -5.006751537322998 + ], + "logprob_reward": -0.006751582026481628, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751582026481628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23228545486927032, + 0.04724135994911194, + 0.14123554527759552, + 0.10202988237142563, + 0.10957203060388565, + 0.0694383904337883, + 0.04785717651247978, + 0.0727410838007927, + 0.12523236870765686, + 0.05236673355102539 + ], + "credit_map": [ + 0.028970595449209213, + 0.031274933367967606, + 0.026423517614603043, + 0.038958918303251266, + 0.02401273511350155, + 0.03097446821630001, + 0.027609549462795258, + 0.02406071312725544, + 0.00874686986207962, + 0.005133847706019878, + 0.0060315923765301704, + 0.007059846539050341, + 0.005021082237362862, + 0.003979706671088934, + 0.0066712601110339165, + 0.004597151186317205, + 0.01634034514427185, + 0.015742335468530655, + 0.016017066314816475, + 0.018701013177633286, + 0.018600642681121826, + 0.012697378173470497, + 0.0215332992374897, + 0.021603461354970932, + 0.013606653548777103, + 0.009895172901451588, + 0.016353042796254158, + 0.00933693815022707, + 0.012461853213608265, + 0.017293035984039307, + 0.010843797586858273, + 0.012239370495080948, + 0.010696692392230034, + 0.01352792326360941, + 0.014977173879742622, + 0.01238041277974844, + 0.02164134941995144, + 0.008348697796463966, + 0.013897811062633991, + 0.014101957902312279, + 0.008799292147159576, + 0.0065561383962631226, + 0.019659550860524178, + 0.008452999405562878, + 0.006219926290214062, + 0.004393883980810642, + 0.007619301788508892, + 0.007737291045486927, + 0.006385945715010166, + 0.00804442074149847, + 0.007112213876098394, + 0.006040850654244423, + 0.00561967957764864, + 0.002314936835318804, + 0.006128022447228432, + 0.006211104337126017, + 0.007835960015654564, + 0.012581304647028446, + 0.008265836164355278, + 0.010261042974889278, + 0.008512313477694988, + 0.007136530242860317, + 0.008477799594402313, + 0.009670288302004337, + 0.008468471467494965, + 0.009547618217766285, + 0.005323040299117565, + 0.010042696259915829, + 0.013047043234109879, + 0.037159912288188934, + 0.009013607166707516, + 0.03262997046113014, + 0.004505825228989124, + 0.008524526841938496, + 0.00875470694154501, + 0.0064499578438699245, + 0.0072099934332072735, + 0.007450115866959095, + 0.005316742230206728, + 0.0041548567824065685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.006751582026481628 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257826764136553 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257826764136553, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257826764136553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1876329928636551, + 0.17653776705265045, + 0.16062381863594055, + 0.06671936810016632, + 0.07056143879890442, + 0.13445381820201874, + 0.04060462489724159, + 0.032276373356580734, + 0.08059344440698624, + 0.04999636113643646 + ], + "credit_map": [ + 0.02647566981613636, + 0.018095064908266068, + 0.034248460084199905, + 0.036411479115486145, + 0.014623532071709633, + 0.012912414036691189, + 0.02683657594025135, + 0.01802978664636612, + 0.03740967810153961, + 0.01880595088005066, + 0.027999361976981163, + 0.020206168293952942, + 0.016947614029049873, + 0.016106126829981804, + 0.023501964285969734, + 0.015560890547931194, + 0.03152342513203621, + 0.02252529188990593, + 0.02304595336318016, + 0.019697614014148712, + 0.016512645408511162, + 0.015865303575992584, + 0.015186886303126812, + 0.016266698017716408, + 0.014976861886680126, + 0.0059767188504338264, + 0.009915833361446857, + 0.006524822674691677, + 0.004352489020675421, + 0.010435120202600956, + 0.005843217018991709, + 0.008694303222000599, + 0.008215179666876793, + 0.009023036807775497, + 0.008546398021280766, + 0.007719309069216251, + 0.01181437075138092, + 0.006330057978630066, + 0.008470128290355206, + 0.010442961007356644, + 0.013804751448333263, + 0.015562760643661022, + 0.02774026058614254, + 0.018048731610178947, + 0.01705074869096279, + 0.014045121148228645, + 0.015054593794047832, + 0.013146849349141121, + 0.00440083397552371, + 0.005471627227962017, + 0.007456135470420122, + 0.004657858982682228, + 0.005756016355007887, + 0.002607879461720586, + 0.006099064834415913, + 0.0041552092880010605, + 0.0033601527102291584, + 0.0038068557623773813, + 0.00398067245259881, + 0.003722647437825799, + 0.003005665261298418, + 0.006719954777508974, + 0.003102012211456895, + 0.004578412976115942, + 0.006868086755275726, + 0.009279090911149979, + 0.004474242217838764, + 0.008665218017995358, + 0.007997103966772556, + 0.019235409796237946, + 0.005706370808184147, + 0.018367933109402657, + 0.00335173518396914, + 0.005848471075296402, + 0.007765895687043667, + 0.00424102321267128, + 0.004060598090291023, + 0.013720818795263767, + 0.004342570900917053, + 0.006665251217782497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005257826764136553 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809867799282074 + ], + "opposite_logprob_tokens": [ + -3.2880985736846924 + ], + "logprob_reward": -0.03809867799282074, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809867799282074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24159827828407288, + 0.0991194099187851, + 0.14595918357372284, + 0.043820105493068695, + 0.11954373121261597, + 0.03150786831974983, + 0.09859013557434082, + 0.022759009152650833, + 0.13640658557415009, + 0.06069573387503624 + ], + "credit_map": [ + 0.02634231187403202, + 0.027630174532532692, + 0.028344450518488884, + 0.04421086981892586, + 0.017765438184142113, + 0.03762288764119148, + 0.030459459871053696, + 0.029222656041383743, + 0.01579180732369423, + 0.012722629122436047, + 0.015876594930887222, + 0.012939873151481152, + 0.009406456723809242, + 0.0056962971575558186, + 0.019191967323422432, + 0.007493771146982908, + 0.02349117584526539, + 0.024859583005309105, + 0.02074924111366272, + 0.018247731029987335, + 0.015767162665724754, + 0.013959281146526337, + 0.014428261667490005, + 0.014456729404628277, + 0.009925409220159054, + 0.004470911342650652, + 0.008914926089346409, + 0.003762194188311696, + 0.0041763740591704845, + 0.004521827679127455, + 0.004251339938491583, + 0.0037971201818436384, + 0.014824925921857357, + 0.015609154477715492, + 0.019432496279478073, + 0.011728620156645775, + 0.019310928881168365, + 0.009557398036122322, + 0.014213780872523785, + 0.014866412617266178, + 0.0032451325096189976, + 0.004686971195042133, + 0.005656888242810965, + 0.004446816630661488, + 0.0034954200964421034, + 0.0032509518787264824, + 0.0033626542426645756, + 0.003363030729815364, + 0.009226770140230656, + 0.016912901774048805, + 0.01158909872174263, + 0.017438126727938652, + 0.012103583663702011, + 0.007013632450252771, + 0.015058299526572227, + 0.009247716516256332, + 0.002224783878773451, + 0.0020467049907892942, + 0.0020679431036114693, + 0.0024165580980479717, + 0.002009131945669651, + 0.005736119579523802, + 0.0024770996533334255, + 0.003780665807425976, + 0.009234459139406681, + 0.010718989185988903, + 0.0058840918354690075, + 0.012184111401438713, + 0.011895750649273396, + 0.040541429072618484, + 0.009239223785698414, + 0.0367085263133049, + 0.004568417556583881, + 0.008930206298828125, + 0.0080980584025383, + 0.00628079567104578, + 0.006535480264574289, + 0.014014670625329018, + 0.005706533789634705, + 0.0065615675412118435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.03809867799282074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761674642562866 + ], + "opposite_logprob_tokens": [ + -0.8261674642562866 + ], + "logprob_reward": -0.5761674642562866, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761674642562866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19662834703922272, + 0.09483829140663147, + 0.09645873308181763, + 0.08770465850830078, + 0.06464117020368576, + 0.019279344007372856, + 0.12041046470403671, + 0.056563280522823334, + 0.21820874512195587, + 0.04526696354150772 + ], + "credit_map": [ + 0.02843576669692993, + 0.01949154958128929, + 0.028506414964795113, + 0.02072884514927864, + 0.01609567366540432, + 0.029790271073579788, + 0.02773333713412285, + 0.02584647759795189, + 0.016241291537880898, + 0.010852314531803131, + 0.0126112075522542, + 0.014114407822489738, + 0.009788709692656994, + 0.009072297252714634, + 0.013358810916543007, + 0.008799253962934017, + 0.013463398441672325, + 0.0166483037173748, + 0.016214020550251007, + 0.012970455922186375, + 0.01199701614677906, + 0.0072739021852612495, + 0.009214910678565502, + 0.008676722645759583, + 0.016712481155991554, + 0.006135146599262953, + 0.016421161592006683, + 0.008239489048719406, + 0.006867121439427137, + 0.007763358298689127, + 0.006732411682605743, + 0.018833493813872337, + 0.008807634934782982, + 0.0095167625695467, + 0.008025790564715862, + 0.007592842448502779, + 0.007619675248861313, + 0.0055289557203650475, + 0.007401375100016594, + 0.010148130357265472, + 0.0024142726324498653, + 0.0020654434338212013, + 0.0035424341913312674, + 0.0024029260966926813, + 0.0026357488241046667, + 0.0017500942340120673, + 0.0022925357334315777, + 0.002175890374928713, + 0.010810118168592453, + 0.024193815886974335, + 0.01511756144464016, + 0.01681484654545784, + 0.01813361980021, + 0.005466863978654146, + 0.016398942098021507, + 0.013474696315824986, + 0.007146415300667286, + 0.007983536459505558, + 0.007057746406644583, + 0.006932391785085201, + 0.005814333911985159, + 0.004946246277540922, + 0.00762296374887228, + 0.009059643372893333, + 0.02028350532054901, + 0.04062476381659508, + 0.019627735018730164, + 0.03102564811706543, + 0.03145035728812218, + 0.02949034981429577, + 0.023171881213784218, + 0.022534506395459175, + 0.0034032026305794716, + 0.0048308540135622025, + 0.007631117943674326, + 0.005910055246204138, + 0.004321228247135878, + 0.004885456524789333, + 0.0046046036295592785, + 0.009680444374680519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.5761674642562866 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "still, vintage film-like close-up captures the face of a majestic Siberian tiger with unwavering precision. In the pale, soft, diffuse light of a winter afternoon, its dense, orange-white fur is sharply outlined by deep black stripes. Its deep, amber eyes exude a primal alertness and majesty, while a few white whiskers, stained with tiny ice crystals, quiver softly on either side of its damp, black nose. The snow-capped coniferous forest in the background is completely blurred by the shallow depth of field, with only the faint outlines of the trees visible. The tiger exhales, solemnly and calmly, slowly and deeply. The warm air instantly condenses into a visible, swirling white mist in the chill air, which is then gently blown away by the cold wind. The entire picture has obvious film grain and the color saturation is slightly reduced, creating a quiet, solemn and wild and poetic atmosphere. In the end, this shot full of power and life slowly fades out to pure black.", + "seed": 28811, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_9/index_1992_seed_28811_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1881665514782071, + "vlm_logprob": -0.1881665514782071, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.1881665514782071, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010839156806468964 + ], + "opposite_logprob_tokens": [ + -6.885838985443115 + ], + "logprob_reward": -0.010839156806468964, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.010839156806468964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20158985257148743, + 0.12334132194519043, + 0.16972693800926208, + 0.05911595746874809, + 0.07876463234424591, + 0.10676953941583633, + 0.058829400688409805, + 0.06321921199560165, + 0.07760140299797058, + 0.06104179471731186 + ], + "credit_map": [ + 0.019591839984059334, + 0.019604748114943504, + 0.02763141505420208, + 0.021508151665329933, + 0.020746693015098572, + 0.019053032621741295, + 0.027415988966822624, + 0.04603796824812889, + 0.017687827348709106, + 0.011751001700758934, + 0.01505205687135458, + 0.008450880646705627, + 0.01424731407314539, + 0.027263665571808815, + 0.01499657891690731, + 0.013891990296542645, + 0.01712452806532383, + 0.017236366868019104, + 0.02078993245959282, + 0.021357566118240356, + 0.022149518132209778, + 0.029193993657827377, + 0.02416430041193962, + 0.0177107322961092, + 0.010505149140954018, + 0.005255589261651039, + 0.006330218631774187, + 0.005530762020498514, + 0.00881149247288704, + 0.009122452698647976, + 0.007421756628900766, + 0.006138532422482967, + 0.010643420740962029, + 0.008949232287704945, + 0.0055534569546580315, + 0.006977065932005644, + 0.011150022968649864, + 0.014135199598968029, + 0.011880950070917606, + 0.009475279599428177, + 0.011665483936667442, + 0.010843507014214993, + 0.01074529904872179, + 0.01263849250972271, + 0.011494074016809464, + 0.01880001649260521, + 0.019582701846957207, + 0.010999957099556923, + 0.006206825375556946, + 0.0040049003437161446, + 0.006767645478248596, + 0.006636090576648712, + 0.005010091699659824, + 0.013608059845864773, + 0.009386674501001835, + 0.007209105882793665, + 0.0055345576256513596, + 0.004413182847201824, + 0.005449831485748291, + 0.006017073057591915, + 0.005587600637227297, + 0.01843021623790264, + 0.007886132225394249, + 0.00990061741322279, + 0.009828362613916397, + 0.005387944635003805, + 0.007014315575361252, + 0.00888090766966343, + 0.007237532641738653, + 0.016456276178359985, + 0.0081489821895957, + 0.01464708149433136, + 0.008046629838645458, + 0.006331244017928839, + 0.0073159364983439445, + 0.007110204081982374, + 0.006519445218145847, + 0.012404042296111584, + 0.006357509642839432, + 0.006956781726330519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.010839156806468964 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008657186292111874 + ], + "opposite_logprob_tokens": [ + -4.758656978607178 + ], + "logprob_reward": -0.008657186292111874, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008657186292111874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3386598229408264, + 0.1855587512254715, + 0.05023882910609245, + 0.03798346593976021, + 0.06632137298583984, + 0.01407746784389019, + 0.0425565280020237, + 0.06272242218255997, + 0.04169897735118866, + 0.1601824313402176 + ], + "credit_map": [ + 0.05247405916452408, + 0.030636383220553398, + 0.045086223632097244, + 0.029602162539958954, + 0.03111361712217331, + 0.05243157222867012, + 0.03688831254839897, + 0.0604274719953537, + 0.0214886125177145, + 0.01439646165817976, + 0.027240661904215813, + 0.019014956429600716, + 0.018912380561232567, + 0.030743880197405815, + 0.028727572411298752, + 0.025034228339791298, + 0.005299185868352652, + 0.00472172861918807, + 0.0057849581353366375, + 0.005644666031002998, + 0.00491471029818058, + 0.010984606109559536, + 0.005553214345127344, + 0.0073357620276510715, + 0.005737056955695152, + 0.003608303377404809, + 0.004015403334051371, + 0.0036492517683655024, + 0.004906494170427322, + 0.004985233768820763, + 0.007158427033573389, + 0.003923298325389624, + 0.006735872942954302, + 0.007638297975063324, + 0.005312444642186165, + 0.006137094460427761, + 0.008171901106834412, + 0.012415518052875996, + 0.011034597642719746, + 0.008875644765794277, + 0.0017945964355021715, + 0.0012586259981617332, + 0.0011081300908699632, + 0.0016656999941915274, + 0.0017629674402996898, + 0.0022419439628720284, + 0.002623804612085223, + 0.001621699775569141, + 0.005936762783676386, + 0.00372232124209404, + 0.004686735104769468, + 0.004804049152880907, + 0.004638736601918936, + 0.006366182118654251, + 0.00769039336591959, + 0.004711346700787544, + 0.00718831829726696, + 0.004232087172567844, + 0.005475255660712719, + 0.006068108603358269, + 0.005443593952804804, + 0.016863759607076645, + 0.008035012520849705, + 0.009416285902261734, + 0.004136241041123867, + 0.002351021859794855, + 0.0037679376546293497, + 0.004213131032884121, + 0.004957238677889109, + 0.011419150047004223, + 0.004341053310781717, + 0.006513203494250774, + 0.01748681627213955, + 0.013897311873733997, + 0.01800108142197132, + 0.018717695027589798, + 0.01324245985597372, + 0.035371746867895126, + 0.023116450756788254, + 0.020348867401480675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.008657186292111874 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026364656165242195 + ], + "opposite_logprob_tokens": [ + -3.65136456489563 + ], + "logprob_reward": -0.026364656165242195, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026364656165242195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30159899592399597, + 0.22254040837287903, + 0.025684436783194542, + 0.020169265568256378, + 0.0434938445687294, + 0.023857012391090393, + 0.03939638286828995, + 0.215138778090477, + 0.0256496574729681, + 0.08247123658657074 + ], + "credit_map": [ + 0.04922642558813095, + 0.02546587400138378, + 0.04009416326880455, + 0.03053997829556465, + 0.030936943367123604, + 0.051321495324373245, + 0.03359048068523407, + 0.04042365401983261, + 0.02085699699819088, + 0.014787287451326847, + 0.03343438729643822, + 0.021215025335550308, + 0.020669154822826385, + 0.05400415509939194, + 0.029689546674489975, + 0.027883851900696754, + 0.0029107537120580673, + 0.0022730082273483276, + 0.003354686312377453, + 0.003811105154454708, + 0.002495964290574193, + 0.003797180950641632, + 0.0032264450564980507, + 0.003815292613580823, + 0.0019204262644052505, + 0.0023722939658910036, + 0.0020251525565981865, + 0.002253884682431817, + 0.0027446835301816463, + 0.003883503843098879, + 0.0026997141540050507, + 0.002269606338813901, + 0.003963533788919449, + 0.003911847248673439, + 0.003515874966979027, + 0.004820808302611113, + 0.005927780643105507, + 0.010226516053080559, + 0.00560982059687376, + 0.00551766250282526, + 0.0033369408920407295, + 0.003549424232915044, + 0.0022116973996162415, + 0.0022705555893480778, + 0.0031662285327911377, + 0.0023827855475246906, + 0.0042846971191465855, + 0.0026546833105385303, + 0.0048134662210941315, + 0.0028773245867341757, + 0.0026056512724608183, + 0.004261551424860954, + 0.004248465411365032, + 0.00885850004851818, + 0.0069426982663571835, + 0.004788723308593035, + 0.02506105601787567, + 0.015382423996925354, + 0.019255273044109344, + 0.02374550886452198, + 0.019098833203315735, + 0.05353589728474617, + 0.028270602226257324, + 0.03078918531537056, + 0.0036122682504355907, + 0.0018852819921448827, + 0.0019844353664666414, + 0.002496379427611828, + 0.0028298962861299515, + 0.005232107359915972, + 0.0036072852090001106, + 0.00400200393050909, + 0.008234214037656784, + 0.006194570101797581, + 0.010259098373353481, + 0.010568360798060894, + 0.007501206826418638, + 0.016054566949605942, + 0.012958458624780178, + 0.01070075761526823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.026364656165242195 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158544182777405 + ], + "opposite_logprob_tokens": [ + -1.7015854120254517 + ], + "logprob_reward": -0.20158544182777405, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158544182777405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21077653765678406, + 0.11337760835886002, + 0.07715492695569992, + 0.055663540959358215, + 0.08810841292142868, + 0.04197880998253822, + 0.05933282524347305, + 0.13934974372386932, + 0.05044907331466675, + 0.16380850970745087 + ], + "credit_map": [ + 0.03459206596016884, + 0.018515927717089653, + 0.029197493568062782, + 0.02096991427242756, + 0.019631201401352882, + 0.03388570621609688, + 0.021893875673413277, + 0.03209036588668823, + 0.012939154170453548, + 0.009211820550262928, + 0.013594725169241428, + 0.011187492869794369, + 0.012936527840793133, + 0.01785755343735218, + 0.02076347917318344, + 0.014886867254972458, + 0.008920560590922832, + 0.0072069852612912655, + 0.008821092545986176, + 0.008687952533364296, + 0.009003331884741783, + 0.015385683625936508, + 0.011117760092020035, + 0.008011572994291782, + 0.005112858023494482, + 0.005776544567197561, + 0.004842272959649563, + 0.007043731398880482, + 0.006442387122660875, + 0.01111961342394352, + 0.009703586809337139, + 0.005622552707791328, + 0.008040204644203186, + 0.009843176230788231, + 0.006940372288227081, + 0.008657857775688171, + 0.009938760660588741, + 0.016322564333677292, + 0.01630857214331627, + 0.012056907638907433, + 0.006479255855083466, + 0.006089579313993454, + 0.0049570775590837, + 0.0037682056427001953, + 0.00481887673959136, + 0.005647412966936827, + 0.0057759941555559635, + 0.00444241426885128, + 0.004956661257892847, + 0.003593324450775981, + 0.006212928332388401, + 0.006103194784373045, + 0.00499179819598794, + 0.015790659934282303, + 0.010234253481030464, + 0.007450007367879152, + 0.012285778298974037, + 0.010739716701209545, + 0.012173004448413849, + 0.01686454936861992, + 0.01312271598726511, + 0.03853082284331322, + 0.017411207780241966, + 0.018221963196992874, + 0.004425001796334982, + 0.0026695560663938522, + 0.0057280408218503, + 0.0052404869347810745, + 0.0064799170941114426, + 0.01171615719795227, + 0.0067482758313417435, + 0.00744163803756237, + 0.02282017469406128, + 0.013909936882555485, + 0.018063673749566078, + 0.019911175593733788, + 0.01328995730727911, + 0.032143332064151764, + 0.025200139731168747, + 0.01847013086080551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.20158544182777405 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933863162994385 + ], + "opposite_logprob_tokens": [ + -0.6933863162994385 + ], + "logprob_reward": -0.6933863162994385, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933863162994385, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2411704808473587, + 0.09237018972635269, + 0.04509850591421127, + 0.05416950210928917, + 0.2354791909456253, + 0.03555676341056824, + 0.04251595586538315, + 0.058121632784605026, + 0.07084807008504868, + 0.12466966360807419 + ], + "credit_map": [ + 0.036772582679986954, + 0.020818039774894714, + 0.025859832763671875, + 0.016741007566452026, + 0.020362872630357742, + 0.04391690343618393, + 0.032941702753305435, + 0.043757546693086624, + 0.00860563199967146, + 0.005943578202277422, + 0.011334329843521118, + 0.008333349600434303, + 0.008650552481412888, + 0.02563009038567543, + 0.011943412013351917, + 0.011929244734346867, + 0.0036048521287739277, + 0.004672787617892027, + 0.004161546938121319, + 0.00461251987144351, + 0.005001798737794161, + 0.011780811473727226, + 0.005586717743426561, + 0.005677476990967989, + 0.009420836344361305, + 0.0049558668397367, + 0.007700515445321798, + 0.005402894224971533, + 0.008092339150607586, + 0.005942295305430889, + 0.006185681093484163, + 0.006469075568020344, + 0.024886325001716614, + 0.0216964203864336, + 0.02067388966679573, + 0.025990484282374382, + 0.031748875975608826, + 0.05635591223835945, + 0.029511667788028717, + 0.02461564540863037, + 0.0038723491597920656, + 0.004350807052105665, + 0.002492026425898075, + 0.00385799678042531, + 0.0038694897666573524, + 0.004769045393913984, + 0.007598420139402151, + 0.0047466326504945755, + 0.0036639634054154158, + 0.003321773372590542, + 0.0023039402440190315, + 0.005006645806133747, + 0.00513243256136775, + 0.010251887142658234, + 0.007984155789017677, + 0.004851160570979118, + 0.00673303846269846, + 0.004706928040832281, + 0.006132027599960566, + 0.006679499987512827, + 0.0051022241823375225, + 0.012787587940692902, + 0.008215714246034622, + 0.0077646165154874325, + 0.006910607684403658, + 0.00375712220557034, + 0.006525243166834116, + 0.007649154402315617, + 0.007558766286820173, + 0.01901298202574253, + 0.008529079146683216, + 0.010905122384428978, + 0.014256509020924568, + 0.010615003295242786, + 0.013438507914543152, + 0.015839392319321632, + 0.010270289145410061, + 0.02703709900379181, + 0.017954545095562935, + 0.015258325263857841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.6933863162994385 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "still, vintage film-like close-up captures the face of a majestic Siberian tiger with unwavering precision. In the pale, soft, diffuse light of a winter afternoon, its dense, orange-white fur is sharply outlined by deep black stripes. Its deep, amber eyes exude a primal alertness and majesty, while a few white whiskers, stained with tiny ice crystals, quiver softly on either side of its damp, black nose. The snow-capped coniferous forest in the background is completely blurred by the shallow depth of field, with only the faint outlines of the trees visible. The tiger exhales, solemnly and calmly, slowly and deeply. The warm air instantly condenses into a visible, swirling white mist in the chill air, which is then gently blown away by the cold wind. The entire picture has obvious film grain and the color saturation is slightly reduced, creating a quiet, solemn and wild and poetic atmosphere. In the end, this shot full of power and life slowly fades out to pure black.", + "seed": 28812, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000170/group_9/index_1992_seed_28812_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15767424385994672, + "vlm_logprob": -0.15767424385994672, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.15767424385994672, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009687081910669804 + ], + "opposite_logprob_tokens": [ + -6.8846869468688965 + ], + "logprob_reward": -0.009687081910669804, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.009687081910669804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7482537627220154, + 0.029727060347795486, + 0.016062844544649124, + 0.015545962378382683, + 0.018802335485816002, + 0.016910668462514877, + 0.03248349577188492, + 0.022720003500580788, + 0.03970370441675186, + 0.059790197759866714 + ], + "credit_map": [ + 0.04039214178919792, + 0.043294504284858704, + 0.3029548227787018, + 0.3064288794994354, + 0.010570712387561798, + 0.014493917115032673, + 0.01952553354203701, + 0.010593300685286522, + 0.0047960299998521805, + 0.005226057022809982, + 0.004163308069109917, + 0.003369362326338887, + 0.002457084134221077, + 0.003625042038038373, + 0.0026510930620133877, + 0.0034390802029520273, + 0.002020095242187381, + 0.0025026367511600256, + 0.0026151039637625217, + 0.003130904631689191, + 0.001122480840422213, + 0.00119735614862293, + 0.001261893194168806, + 0.0022123735398054123, + 0.002211760263890028, + 0.0025930111296474934, + 0.0033286057878285646, + 0.0017828189302235842, + 0.0013281910214573145, + 0.0010823231423273683, + 0.001504818326793611, + 0.0017144326120615005, + 0.004714777693152428, + 0.005110481753945351, + 0.0020396148320287466, + 0.001434431760571897, + 0.0011370256543159485, + 0.0013179483357816935, + 0.0015975742135196924, + 0.0014504806604236364, + 0.003181824693456292, + 0.0024166316725313663, + 0.0022140932269394398, + 0.001750280149281025, + 0.0011438180226832628, + 0.0013970842119306326, + 0.0014753310242667794, + 0.0033316032495349646, + 0.0054716081358492374, + 0.0062392354011535645, + 0.004266980569809675, + 0.005459221079945564, + 0.0027060501743108034, + 0.0023659721482545137, + 0.002631223062053323, + 0.003343202406540513, + 0.0031608992721885443, + 0.0036762957461178303, + 0.0028761497233062983, + 0.0038418713957071304, + 0.0018085883930325508, + 0.0015556326834484935, + 0.002013659104704857, + 0.0037869049701839685, + 0.007597907446324825, + 0.007461810950189829, + 0.004315062426030636, + 0.007062235381454229, + 0.002714555012062192, + 0.0028661410324275494, + 0.002859440166503191, + 0.004826544318348169, + 0.010005229152739048, + 0.012150838971138, + 0.006743886042386293, + 0.009565804153680801, + 0.003126263851299882, + 0.0053222463466227055, + 0.004706600680947304, + 0.008169321343302727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.009687081910669804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007648114114999771 + ], + "opposite_logprob_tokens": [ + -4.88264799118042 + ], + "logprob_reward": -0.007648114114999771, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007648114114999771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6382411122322083, + 0.03389463573694229, + 0.08673366159200668, + 0.021860886365175247, + 0.021018914878368378, + 0.04511541873216629, + 0.01847194880247116, + 0.03556923195719719, + 0.032743584364652634, + 0.06635063141584396 + ], + "credit_map": [ + 0.03494798392057419, + 0.03575129434466362, + 0.25994566082954407, + 0.2575590908527374, + 0.00967901386320591, + 0.014529028907418251, + 0.016502516344189644, + 0.009326480329036713, + 0.005984209012240171, + 0.006240210961550474, + 0.005428141914308071, + 0.006080670282244682, + 0.0025731923524290323, + 0.002612736076116562, + 0.00258614681661129, + 0.0023893294855952263, + 0.01769031211733818, + 0.018371399492025375, + 0.015057061798870564, + 0.013629927299916744, + 0.0048482343554496765, + 0.005409318022429943, + 0.005072886124253273, + 0.006654525641351938, + 0.003423163900151849, + 0.0036233970895409584, + 0.004081488586962223, + 0.002334626391530037, + 0.0017948502209037542, + 0.0018045229371637106, + 0.0022174064069986343, + 0.0025814296677708626, + 0.004599934909492731, + 0.005132630001753569, + 0.0022965106181800365, + 0.002390632638707757, + 0.001522360253147781, + 0.001315517001785338, + 0.0019973190501332283, + 0.0017640111036598682, + 0.007506738416850567, + 0.0060334703885018826, + 0.009620669297873974, + 0.008505171164870262, + 0.0017653197282925248, + 0.0029937955550849438, + 0.0023678988218307495, + 0.0063223568722605705, + 0.003312727203592658, + 0.0028636667411774397, + 0.0022631767205893993, + 0.0034041213802993298, + 0.0016223808052018285, + 0.0015810270560905337, + 0.0014902696711942554, + 0.0019345786422491074, + 0.005112342536449432, + 0.006967814639210701, + 0.004184139426797628, + 0.005646451376378536, + 0.003618791000917554, + 0.0023680769372731447, + 0.0035021023359149694, + 0.00416951160877943, + 0.007103639654815197, + 0.007601450197398663, + 0.0038316496647894382, + 0.004552335478365421, + 0.0017999395495280623, + 0.0021663124207407236, + 0.001972353085875511, + 0.0037159037310630083, + 0.009698251262307167, + 0.007295569870620966, + 0.010362804867327213, + 0.017246153205633163, + 0.00359190721064806, + 0.006506044417619705, + 0.005278104916214943, + 0.006371788214892149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.007648114114999771 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633799687027931 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633799687027931, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633799687027931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6000884771347046, + 0.033813729882240295, + 0.03633620962500572, + 0.02977716736495495, + 0.13092945516109467, + 0.00980703067034483, + 0.08169184625148773, + 0.011041824705898762, + 0.039653047919273376, + 0.02686116099357605 + ], + "credit_map": [ + 0.03348883241415024, + 0.03352198004722595, + 0.24549204111099243, + 0.24380190670490265, + 0.008387886919081211, + 0.011553103104233742, + 0.015266836620867252, + 0.00857590138912201, + 0.006428911350667477, + 0.006615594960749149, + 0.006542595103383064, + 0.0062352074310183525, + 0.001839925767853856, + 0.0017504682764410973, + 0.0023542174603790045, + 0.002046810695901513, + 0.007053114473819733, + 0.007031846325844526, + 0.006043393164873123, + 0.0064651439897716045, + 0.002065210836008191, + 0.002101134741678834, + 0.002731841988861561, + 0.0028445252683013678, + 0.005129100289195776, + 0.0055895899422466755, + 0.0067113954573869705, + 0.0042538754642009735, + 0.0018165562069043517, + 0.001524425344541669, + 0.001958490116521716, + 0.0027937330305576324, + 0.03381595388054848, + 0.03549233078956604, + 0.01782175526022911, + 0.013542946428060532, + 0.006532224360853434, + 0.006523202173411846, + 0.005523539613932371, + 0.011677512899041176, + 0.0019155732588842511, + 0.0020192787051200867, + 0.0013918945332989097, + 0.0013119200011715293, + 0.0005401161033660173, + 0.0007503137458115816, + 0.0007210003095678985, + 0.0011569340713322163, + 0.014916820451617241, + 0.016697095707058907, + 0.009919910691678524, + 0.018564384430646896, + 0.00533105805516243, + 0.003946478944271803, + 0.005565368104726076, + 0.006750732660293579, + 0.0016880945768207312, + 0.0019828323274850845, + 0.0013435883447527885, + 0.002427852712571621, + 0.0007045030943118036, + 0.0006442071753554046, + 0.0008753525326028466, + 0.0013753935927525163, + 0.009067786857485771, + 0.010085155256092548, + 0.004610372707247734, + 0.005268249195069075, + 0.002828712807968259, + 0.0027549618389457464, + 0.00207505002617836, + 0.0029627576004713774, + 0.005093744024634361, + 0.006758581381291151, + 0.0028968886472284794, + 0.0037858160212635994, + 0.0018909955397248268, + 0.0020891453605145216, + 0.0018571849213913083, + 0.002488805213943124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.007633799687027931 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002770900726318 + ], + "opposite_logprob_tokens": [ + -2.6950278282165527 + ], + "logprob_reward": -0.07002770900726318, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002770900726318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16604867577552795, + 0.10930170118808746, + 0.2225719839334488, + 0.03121701441705227, + 0.03043597750365734, + 0.21845269203186035, + 0.018313677981495857, + 0.04602427780628204, + 0.031003186479210854, + 0.12663079798221588 + ], + "credit_map": [ + 0.015361208468675613, + 0.014264731667935848, + 0.05528474971652031, + 0.057463932782411575, + 0.0049507468938827515, + 0.006443483754992485, + 0.005973947700113058, + 0.006305891089141369, + 0.016419420018792152, + 0.0134813217446208, + 0.029642419889569283, + 0.02407038025557995, + 0.006543065886944532, + 0.006003746762871742, + 0.006054690107703209, + 0.007086664903908968, + 0.029255691915750504, + 0.033474430441856384, + 0.050694432109594345, + 0.06160563975572586, + 0.010045950300991535, + 0.010265935212373734, + 0.011892353184521198, + 0.015337567776441574, + 0.006130923982709646, + 0.0038780211471021175, + 0.007022647652775049, + 0.005798290949314833, + 0.0017622736049816012, + 0.001988611649721861, + 0.0017999991541728377, + 0.0028362504672259092, + 0.005545672029256821, + 0.0059087201952934265, + 0.0055453418754041195, + 0.0033448669128119946, + 0.0020345274824649096, + 0.0022083288058638573, + 0.0034903171472251415, + 0.002358206082135439, + 0.04427684471011162, + 0.038462162017822266, + 0.04287626966834068, + 0.03992876783013344, + 0.010541332885622978, + 0.012250742875039577, + 0.011278047226369381, + 0.01883852295577526, + 0.003514247015118599, + 0.003263241844251752, + 0.0019400152377784252, + 0.0026463018730282784, + 0.002079558093100786, + 0.0017389672575518489, + 0.0013923905789852142, + 0.0017389568965882063, + 0.005985160358250141, + 0.00558245275169611, + 0.006252444349229336, + 0.008803372271358967, + 0.003347342601045966, + 0.00284904008731246, + 0.004189125262200832, + 0.009015346877276897, + 0.006107468623667955, + 0.007311325054615736, + 0.003447418799623847, + 0.005099987145513296, + 0.0022059159819036722, + 0.0025470457039773464, + 0.002499338472262025, + 0.0017846861155703664, + 0.017412366345524788, + 0.010990997776389122, + 0.021108074113726616, + 0.03798874095082283, + 0.006401679012924433, + 0.009939957410097122, + 0.011384665034711361, + 0.01140433456748724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.07002770900726318 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a Siberian tiger's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the tiger have dense orange-white fur with deep black stripes and white whiskers stained with ice crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a snow-capped coniferous forest that is completely blurred by shallow depth of field?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the tiger exhale slowly, creating a visible swirling white mist that is then blown away by wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video exhibit a vintage film-like style with obvious film grain and reduced color saturation?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.693374514579773 + ], + "opposite_logprob_tokens": [ + -0.693374514579773 + ], + "logprob_reward": -0.693374514579773, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.693374514579773, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.7049076557159424, + 0.02011951431632042, + 0.0552072636783123, + 0.04142177850008011, + 0.014531542547047138, + 0.025870230048894882, + 0.03520987182855606, + 0.030941147357225418, + 0.04281771555542946, + 0.028973259031772614 + ], + "credit_map": [ + 0.051338501274585724, + 0.04958884418010712, + 0.2700424790382385, + 0.2796190083026886, + 0.010937441140413284, + 0.01349365059286356, + 0.01946611888706684, + 0.010421671904623508, + 0.003031675238162279, + 0.0028667058795690536, + 0.003035777248442173, + 0.003322308650240302, + 0.0019362857565283775, + 0.0022270926274359226, + 0.0016747163608670235, + 0.0020249520894140005, + 0.009938719682395458, + 0.011354773305356503, + 0.009405884891748428, + 0.009649666957557201, + 0.003297065384685993, + 0.003468356793746352, + 0.003678712295368314, + 0.004414085764437914, + 0.007135939784348011, + 0.00631605414673686, + 0.007746514864265919, + 0.0050625004805624485, + 0.0035662399604916573, + 0.0034617395140230656, + 0.0037164518143981695, + 0.004416338168084621, + 0.0025227286387234926, + 0.0029926339630037546, + 0.002252620179206133, + 0.0014100918779149652, + 0.0011171509977430105, + 0.0013636614894494414, + 0.0011274799471721053, + 0.0017451762687414885, + 0.0043059540912508965, + 0.0040056901052594185, + 0.0030484339222311974, + 0.005029739812016487, + 0.001553280744701624, + 0.0016703095752745867, + 0.002175046131014824, + 0.004081774037331343, + 0.006462614517658949, + 0.005425882060080767, + 0.004435285925865173, + 0.007797398138791323, + 0.002021683845669031, + 0.002313115866854787, + 0.002915374469012022, + 0.003838515607640147, + 0.004906956106424332, + 0.005408674478530884, + 0.003721001325175166, + 0.006630596239119768, + 0.0020612466614693403, + 0.0017815371975302696, + 0.002394744660705328, + 0.004036389756947756, + 0.007880663499236107, + 0.00490051181986928, + 0.005388602614402771, + 0.009852960705757141, + 0.002388229826465249, + 0.003270838875323534, + 0.003737545106559992, + 0.005398362874984741, + 0.005093371029943228, + 0.007400724571198225, + 0.0031370953656733036, + 0.004480138421058655, + 0.0018143273191526532, + 0.002259213011711836, + 0.002332608914002776, + 0.0024557795841246843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.693374514579773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file